AI models and tools worth knowing about

Everything the AI channels we follow have named, each with a plain description of what it is, how you get it and what it needs to run.
362 items from 32 videos · ✓ marks the ones we have tried ourselves. · RSS feed · rebuilt 2026-11-02

2026-11-01 AI Search

Unlimited OCR 2

Baidu's second-generation document reader, which pulls text out of documents of any length by keeping a fixed-size set of references as it slides through the pages.

open source, mit · needs the published tests used an NVIDIA A800 80GB

UniAction

UniAction represents any action as a sparse set of action elements, letting force, camera control, hand pose and robot pose condition one world model.

StreamVibe

StreamVibe studies proactive streaming video agents that infer implicit user intents, track evolving requests, and decide when and how to assist.

Spark-X2.6-1.7B

A 1.7-billion-parameter chat model built to run on your own device, with three selectable amounts of thinking to trade quality against speed.

open weights, apache-2.0

SLEEP

The sleep framework: A new way for LLMs to retain knowledge in continual learning Skip to main content The sleep framework: A new way for LLMs to retain knowledge in continual learning October 27, 202.

SenseNova-Loop 9B

A model that reads text and images together and runs its 32 layers 56 times over in a loop, so roughly 9 billion stored weights do the work of a much bigger model.

open weights, apache-2.0 · needs the published deployment examples use 4 to 8 GPUs

SenseNova 6.8 Code2Video

A 27-billion-parameter model that makes a video by writing animation code and running it, instead of generating the picture frames directly. The weights are Apache-2.0, but the pipeline is built around Claude Code as the recommended driver, which is a separate paid service.

open weights, apache-2.0 (the weights on Hugging Face carry the licence; the GitHub repo itself has no licence file)

SceneBloom

SceneBloom compiles diffusion-sampled furniture arrangements into collision-free, physically stable 3D rooms that train better object-navigation agents.

PhysAg3D

PhysAg3D couples 3D generation with agentic reasoning to turn a single image into detailed, image-aligned 3D assets with plausible interactions.

Paradee 1.2

Paradee now speaks four voices without growing in size or losing quality.

OverSeer

OverSeer is a multi-agent video world model that augments the agents.

OpenAI scientific discoveries

Open-H3-Regenerate-2K

An open stand-in for MiniMax's closed 2K upscaling step: it takes a low-resolution H3 video draft and redraws it at up to 4K in three passes. Licensing is a patchwork rather than one blanket grant — the main refiner follows MiniMax's own community licence, and the borrowed parts each keep theirs.

open weights, MiniMax H3 community licence for the refiner, other components under their own terms · needs the published timings used a single NVIDIA B200: about 22 seconds for a 5-second clip at 2K

NVIDIA SimReady Assets

A library of 2,677 ready-to-simulate 3D objects in USD format — furniture, medical kit, car parts, food — for dropping into physics simulations. About 490 GB in total.

open source, CC-BY-4.0

NVIDIA Nemotron 3.5 Transcribe

NVIDIA's speech model: it transcribes and translates across 40 languages, labels who is speaking, and will answer questions about a recording up to an hour long. 34B parameters in total, about 4B used per token.

open weights, OpenMDW-1.1, commercial use permitted · needs one H100 80GB is the stated deployment target; also runs on A100, H200, L40S and Blackwell cards

Nanbeige4.3-33B

A 33-billion-parameter model aimed at agentic coding, which its authors built mainly as the teacher for the smaller models they plan to distil from it.

open weights, apache-2.0

MUNDO

Builds an editable 3D version of a real place from a handful of photos of it, so you can move the camera around, change the objects and test a robot in it. Code and weights are not out yet — the repository says end of November 2026.

research project

MotionBricks

Generates human and robot movement on the fly from a library of over 350,000 motion skills, so a character can walk, navigate a room and handle objects without hand-animated clips.

open source, Apache-2.0 code; weights under the NVIDIA Open Model Agreement, which permits commercial use

MatAnyone-E

A faster way to cut a moving subject out of high-resolution video, hair and soft edges included, by carrying less of the previous frame forward. The project page presents it as released, but the code repository still lists the inference code and model files as to-do, so there is nothing to download yet.

research project · needs the published tests used 3.1 GB of GPU memory at 4K, on an H200 and also on an RTX 3090

LFM3

Liquid AI's new family of six small models — text, vision, audio, omni and decision-making — all around 2B parameters and built to run on a phone or laptop rather than a server. The 1.7B text model fits in about 1 GB at 4-bit.

open weights, LFM Open License v1.0 · needs about 1 GB at 4-bit for the 1.7B text model; a full 128K-token context adds 1.6 GB of cache

IDOL

Turns a single photo into a 3D object made of Gaussian splats, filling in the sides the photo never showed, at a quality level you choose.

open source, Apache-2.0 · needs under 8 GB of GPU memory; the published tests ran it on cards down to an RTX 3080 10GB, at 46 seconds per image

Hy-SimulMT

Tencent's live Chinese-English interpreter: it starts translating while you are still speaking and speaks the result back in a copy of your own voice.

open source, apache-2.0 · needs an NVIDIA GPU on Linux; the 30B version needs about 57 GiB at BF16, or one H200, with FP8 and GGUF builds offered for smaller cards

Groot

A family of open-weight reasoning models, from 19B up to 280B parameters, that read images as well as text and are tuned to answer in fewer words than similar-sized rivals.

open weights, apache-2.0 · needs for the 19B version: a single 48GB or 80GB GPU at BF16, or the Q4 GGUF build at 128K context on a 16GB RTX 4060 Ti and 256K on a 24GB RTX 4090

GPT-6.1 Terra

OpenAI's mid-priced model, aimed at office tasks and at driving a computer on your behalf. Its selling point is cost: the API price is half that of the model it replaces.

paid service · needs none, runs in the cloud

Genie 4

A DeepMind model that invents an interactive 3D scene, with sound, that you can move through and change by typing instructions at it.

research project, a preview for trusted testers with a waitlist; no weights released · needs none, runs in the cloud

Gemini 4 Carbon

A Gemini model Google has specialised for security work — finding vulnerabilities in code, checking whether they are actually exploitable, and writing patches. The video presented it as a launch, but Google says it is going first to a selected set of cyber defenders and partners, not to the public.

paid service, invite-only at launch through Google's Fairwind Program; Google states $2 per million input tokens and $10 per million output tokens for when it opens up · needs none, runs in the cloud

Clover 35B

A 35-billion-parameter language model that writes whole blocks of text at once by repeatedly refining them, rather than one word after another; about 3 billion parameters are used per token.

open weights, apache-2.0 · needs the vendor reports it running at INT4 on one 16GB NVIDIA RTX 5070 Ti

AutoSwarm

A coordinator that hands a job to whichever coding agents you already have installed, choosing them by what they are good at and passing context between them. It supplies no model of its own — you still need accounts for the agents it drives.

open source, mit

Audio8 TTS Realistic 0.6B

A small text-to-speech model, 0.6 billion parameters, that reads text aloud in 23 languages and can copy a voice from one short sample. Unlike the earlier Audio8 release, this one is plainly Apache-2.0.

open weights, apache-2.0 · needs a CUDA-capable GPU is recommended; the published tests used an NVIDIA H20 at BF16

Apex 35B

A general-purpose text model from Apex Intelligence, pitched at research and scientific work, which the company says it built using automated AI research rather than by hand. It is a mixture-of-experts: 35B parameters in total but only 3B used per token.

open weights, Apache-2.0 · needs the BF16 weights are about 72 GB; one 141 GB GPU (H200, B200 or H20-3e) runs it on its own, smaller cards need splitting across several

Anthropic S-1

Anthropic's IPO paperwork, filed with the US securities regulator on 26 October 2026: it registers Class A shares for a planned Nasdaq listing under the ticker ANT and discloses the company's finances. The filing is not yet effective, so no shares can be sold on it.

news article

2026-10-28 Mickmumpitz

🔗 Learn more about NVIDIA RTX AI

The video's sponsor link rather than a release: NVIDIA's marketing page for GeForce RTX PCs, arguing you should run AI locally on an RTX GPU and listing its own tools (ChatRTX, NIM microservices, AI blueprints) plus integrations with ComfyUI, Blender, Adobe and DaVinci Resolve. The page names no prices, and it does not mention the 'RTX video nodes' the video credits for its 4K upscaling.

unclear, a marketing page for graphics cards rather than anything you can download or subscribe to · needs an NVIDIA GeForce RTX GPU; the page lists the RTX 50 and 40 series with their AI TOPS figures but states no minimum model

2026-10-25 AI Search

WeVisOne-VL

Tencent's image-understanding model that does detection, counting, pointing, text recognition, segmentation and cut-out mattes in one model rather than several. Released in 2B and 4B sizes.

open weights; code under Apache-2.0

Sky-001

Skyfall AI's design model for engineering: from a text prompt or a 2D sketch it plans an object as an ordered sequence of CAD or BIM construction steps, so each part can be rebuilt in normal engineering software. Every speed and cost figure is the company's own.

unavailable; an initial research preview with a waitlist, full commercial release targeted for early 2027

Sage-A1.5

Shanghai AI Lab's agentic model for long scientific and engineering tasks: it can split work across sub-agents and pause mid-task to review what has worked and change course. 35B total parameters with about 3B active, built on Qwen3.6-35B-A3B.

open weights, Apache-2.0

PURE

A training method that sharpens existing video-to-3D reconstruction models using ordinary unlabelled video, so no hand-made 3D ground truth is needed; it reconstructs depth, camera motion and 3D points from a clip. An academic paper (ACM MM 2026) from HKUST, with code and fine-tuned weights released for the pi3, da3 and VGGT models.

open source

OpenAI research agents

OpenAI's write-up of a research collaboration in which thousands of AI agents combed through a large mouse-brain gene-perturbation dataset in about half an hour, turning up leads for Alzheimer's and Parkinson's research; it is an account of the work, not something you can use.

research project

MiniMax M3.1 Flash

MiniMax's fast, cheap model for agent work, coding and computer use, priced at $0.15 per million input tokens and $0.60 per million output. MiniMax says the weights will be released later, so for now it is API-only.

paid service · needs none, runs in the cloud

Ming 3.0 TTS

Ant Group's text-to-speech model: it reads a script aloud, lets you describe the voice and delivery you want in plain words, clones a voice from a sample, and covers seven languages. Released in 0.5B and 1.3B sizes.

open source, MIT

MemT 1.0

A text model that writes what matters into a small internal memory state, so it can still answer about something after that part of the conversation has dropped out of its context window. 36B total parameters with 3B active, built on a Qwen3.6-35B-A3B backbone; a much larger MemT-1.0-Flash sibling also exists.

open weights, Apache-2.0

Ling 3.0 Flash Santé

A healthcare-tuned version of Ling-3.0-flash for medical reasoning and research, 124B total parameters with only 5.1B active per token. Its own model card warns it is not a medical device and its answers may be wrong.

open weights, MIT · needs deployment examples use four 141 GB GPUs or eight 80 GB H100/H800-class GPUs

Kimi K3.1 Pro

Moonshot's new flagship chat and coding model, described as 3T-class with a 1-million-token context window, plus a Highspeed variant that runs 2-3x faster. Billed as open-weight, but the blog says the weights themselves only arrive by 4 November 2026 — until then it is API-only.

hosted service for now; open weights promised by 4 November 2026, licence not stated

JST2

Viggle's 3D world model that builds scenes, animation-ready characters and full-body motion from text, images or video, generating a scene in about a second, and can reconstruct real places from a few views.

free to use in the browser via Viggle's PINOC app; commercial use goes through the paid API (under $0.01 per scene) · needs none for users; Viggle runs inference on NVIDIA B200

Ideogram Vector

A tool that generates editable SVG vector artwork — logos, print designs, illustrations — from a text prompt or an uploaded image, with individual shapes you can still adjust.

paid service, charged per image ($0.13-$0.32 per text-to-vector SVG, $0.10 per image conversion) with a free first vector · needs none, runs in the cloud

Hy-Nomas

Tencent's interactive video world model: instead of producing one fixed clip, it generates 720x1280 video a segment at a time while you steer it live with new text prompts and camera moves, keeping earlier scenery and objects consistent via a persistent cache.

open weights, Apache 2.0 · needs eight GPUs for the standard setup; a low-memory INT8 build runs on a single RTX 4090 with CPU offload

Hi3D 1.0

An image-to-3D model, previously called Sparc3D, that turns a single picture into a 3D asset with geometry and texture. Trained on roughly 2 million curated 3D assets, with code, weights and an online demo released.

open source

Fibo 2.0

An 8B image model from Bria that both generates pictures from text and edits existing ones in a single model, including masked region edits and combining up to four reference images.

open weights, CC BY-NC 4.0 — non-commercial only, commercial use needs a paid agreement with Bria · needs CUDA GPU; tuned for H100-class Hopper GPUs with FlashAttention-3

FastH3 Omni

A slimmed-down (distilled) version of MiniMax H3 that makes reference-guided video-with-sound generation much faster: you give it a prompt plus images, clips or a voice sample and it returns a finished video. The page claims 5.6-11.9x lower latency than the base model, so the video's "12x" is the very top of that range rather than the typical figure.

open weights · needs benchmarked on eight GB200 GPUs; NVFP4 quantised weights are offered for DGX Spark or an RTX 5090

Dessa

A speech-in, speech-out conversational model from a research lab inside Block, trained only on recordings of people talking rather than on text, so it can listen and speak at the same time and carry emotion and timing. Not yet released — the page offers only a waitlist and early access for safety researchers.

research project, waitlist only · needs stated to run locally on a MacBook Pro M4 Max or better

Claude Duet

An Anthropic phone and web app where Claude works in the background on your behalf — sending a morning brief, watching your connected apps and doing tasks you hand it, asking permission before anything consequential like sending a message or making a payment.

paid service, research preview by waitlist for US Pro and Max subscribers only · needs none, runs in the cloud; iOS, Android or web

Carbon-X

A 14.8B language model that reads raw DNA sequences inside a normal chat and explains them in English — for example judging whether a mutation is harmful, or spotting gene-controlling regions. It averages 66.7% accuracy across five DNA benchmarks.

open weights, Apache 2.0 · needs fits on a single 80 GB GPU in bfloat16

Bark3

A small English speech-to-text model (0.6B parameters, a 163 MB download) that transcribes audio on your own machine rather than in the cloud. It is a fine-tune of NVIDIA's Parakeet, not a model built from scratch.

open weights, CC-BY-4.0 · needs runs on CPU, Apple Silicon (macOS 15+) or an NVIDIA GPU with CUDA 12.8; about 1.5s per minute of audio on an M2 laptop

2026-10-23 AI Search

Deepseek V4.1 1017

The 17 October refresh of DeepSeek-V4.1-Flash: same architecture as before, reads images as well as text, 1M-token context. Its own card calls it "a minor update" but reports large jumps on agent and reasoning tests — Terminal-Bench 4.0 from 31.2 to 42.3, HLE from 36.8 to 42.6 — while producing shorter answers for the same work. The card itself states no parameter count; the video's 552-billion/16-billion-active figures come from the earlier V4.1-Flash write-up, and Hugging Face's own file metadata for both repos reports a higher total, so treat any single number for this model as unsettled.

open weights, MIT, on Hugging Face

DeepInfer

DeepSeek's own inference engine — the software that serves its models — published as source. It currently supports exactly one model, DeepSeek-V4.1-Flash, and gives you offline inference. Contrary to the video's framing of it as the stack behind their data centres, the README says the parts that turn this into a live service for millions of users are deliberately not included: the KV-cache system, prefill/decode disaggregation, and load balancing across replicas.

open source, Apache-2.0 · needs NVIDIA SM100-class GPUs or Huawei Ascend 950 NPUs; Linux, CUDA 13 or CANN 9.2, Python 3.12 — data-centre hardware, not a consumer card

2026-10-20 AI Search

MOSS Image

The ComfyUI team's repackaging of OpenMOSS's text-to-image model into single files you can drop straight into ComfyUI, bundled with the Qwen3.5-9B text encoder and the VAE it needs; the original weights are the OpenMOSS team's, not ComfyUI's.

open weights, Apache-2.0

lora.safetensors

A small add-on file for MOSS Image, trained by lodestones, that cuts generation from 50 steps down to 8 so pictures come out several times faster.

open weights, Apache-2.0

Kromoss

A community retrain of MOSS Image's base model by lodestones: a 5B text-to-image model that makes 1024px pictures and ships with its own 8-step fast mode.

open weights, Apache-2.0

ai toolkit

A training tool for teaching image, video and audio models new styles or subjects on consumer hardware; the video uses it to train LoRAs for MOSS Image.

open source, MIT · needs an Nvidia GPU; the project states no specific amount of VRAM

2026-10-18 AI Search

UniForge 3D

Turns a few photos of an object taken from unknown angles into a complete 3D model, reconstructing the visible parts and generating the hidden ones in one shared space; despite the video saying the code and models are out, the project page still lists the code as coming soon and only the arXiv paper is up.

paper only, code listed as coming soon

StreamRender-H3

NVIDIA's real-time game renderer: an ordinary Three.js game engine keeps track of the world and a streaming video model repaints each frame photorealistically as you play.

open weights, code under Apache-2.0 · needs 8 NVIDIA GB200 GPUs for the renderer, with the decoder on a further GB200 node

Streaming Code World Model

Code and LoRA weights for generating video in continuous 0.71-second chunks, each one conditioned on the chunks before it, using a four-step distilled version of MiniMax-H3.

open weights, no licence stated in the repository · needs one NVIDIA H200 with 141 GB; the repo states 80 GB GPUs are not supported

Step 5 Preview

StepFun's flagship Chinese model for long-running agent work, 600 billion parameters in total with 27 billion active at a time and a 1-million-token context; the weights were published this week after an earlier API-only preview.

open weights, also sold as a paid API · needs about 1.2 TB of GPU memory at BF16 (8x H100), or roughly 600 GB at FP8

RP1

A complete open humanoid robot platform: mechanical CAD, circuit schematics, PCB files, bill of materials, firmware, a 500 Hz control stack and training code, in configurations from 24 to 42 joints.

open source, GPL-3.0 · needs a physical robot you build yourself; the hardware designs are published

PingpongDuo

A control system that lets two humanoid robots play table tennis as a doubles pair, alternating strikes and stepping aside so the partner can take the next shot; shown on real robots sustaining rallies of more than ten strikes.

paper only

NVIDIA Nemotron 3.5 Super VL

NVIDIA's open vision-language model for enterprise agents that read documents and screens and call tools, 120 billion parameters with 12 billion active at a time.

open weights, OpenMDW-1.1 · needs 4x H100 or 2x GB200 for the full 1-million-token context

Next-Concept Modeling

A Stanford method that writes text like an outline being filled in: it predicts a coarse sketch of the whole passage, then refines it in a few parallel passes instead of one word at a time, which the paper reports as 10-31x faster than comparable diffusion language models.

open source, MIT

Muse Wonder

Meta's new top-end closed model for agents that hold on to context and keep working towards a goal; sold through Meta's developer API and bundled with the Muse Code subscription.

paid service · needs none, runs in the cloud

Muse Spark Open

Meta's open-weight model for coding and reasoning, 816 billion parameters in total; Meta pitches it as desktop-ready, but the desktop in question is two Apple M5 Ultra Mac Studios with up to 512 GB of memory each.

open weights, Muse Model License with limited commercial terms · needs two Apple M5 Ultra Mac Studios with up to 512 GB unified memory each, per Meta's own setup

Multics

A world model that generates each player's own view of the same shared virtual scene at the same time, keeping everyone's movements and surroundings consistent; a playable demo on the page runs 6, 8 or 10 viewpoints at once.

open weights, Apache-2.0

MOSS-Image

An open text-to-image model you can steer with a JSON layout as well as a prompt, so you can say exactly where each object goes in the picture.

open weights, Apache-2.0 · needs an NVIDIA GPU; the model is 5 billion parameters

MosaiChunk

An add-on memory for AI video generators: it keeps snapshots of what objects looked like earlier in a clip and feeds them back when those objects return to frame, so faces and scenery stop changing appearance after going off screen. Code, weights and a benchmark are released, but the weights only sit on top of someone else's video model, which you have to obtain separately.

open weights, non-commercial only — the image-to-video checkpoint is CC BY-NC-SA 4.0 and the text-to-video one falls under the MiniMax-H3 Community License; the release grants no rights to the underlying video model it needs · needs the published inference setups use one NVIDIA H200 for image-to-video and eight H200s for text-to-video; training in the paper used 16 to 32 H200s

Matrix-Game 4.0

A playable AI game world: an ordinary game engine handles the rules, physics and your input while a fast video model repaints its rough frames into detailed visuals in real time, and a coding agent writes new rules from plain-language requests.

open source, inference and game code on GitHub with scene weights and game assets on Hugging Face under apache-2.0 · needs a single consumer NVIDIA GPU; the headline 60 fps at 480p figures are on an RTX 5090, with optimised builds reported for RTX 5080 and 5070 Ti

Ling-3.1-flash

Despite the "flash" name this is a very large model: 560 billion parameters with 25 billion active per token, a 1M-token context, aimed at software engineering and agent work, and far too big to run at home.

open source, mit · needs the model card's own deployment recipes use 8 x H200 or 4 x GB300 GPUs

LFM2.5-24B-A2B

A 24-billion-parameter reasoning model that only activates 2 billion parameters per token, so it is fast enough to run agents locally on a laptop or workstation instead of in the cloud.

open weights, LFM Open License v1.0: commercial use is free only below US$10,000,000 annual revenue

Krea 3

An image editing model that applies an edit in under eight seconds at 2K while keeping the same product or character recognisable; billed per image, not open weights.

paid service, $0.06 per 2K image and $0.03 at 1K on krea.ai, also resold through fal, Together AI, OpenRouter and Replicate

Kling 4

Kling's commercial web studio for generating video and images from text or reference pictures; version 4 is the current model, sold on steadier motion and matching generated audio.

hosted service, the site quotes no price on its front page

InfiniSplat v2

Rebuilds a 3D scene from a handful of photographs in one pass as a Gaussian splat, and does not need to be told where the cameras were.

open source with a non-standard licence (Project Registration License v1.0): free for academic, educational and personal use, organisations must register first but pay nothing; a Hugging Face demo is also up

Google Quantum

Not an AI release: Google reports that an error-correction technique called magic state cultivation cut the error rate of a key quantum operation about 40-fold on its Willow chip, published as a paper in Science.

paper only

Gemini 4 Argon

Google's new top-end Gemini model, pitched at hard coding, scientific and legal work, and rolling out globally from launch day.

paid service, needs a Google AI Pro or Ultra subscription or the paid API, introductory API rate $2 per million input tokens and $10 per million output

GaussiAnimate

Takes a captured 4D scan of something moving and automatically fits it with a skeleton and joints, so it can then be re-animated rather than only replayed.

research project, the page promises code "for research purposes" but it is not out yet

FLUX 3 Image Klein

The downloadable member of Black Forest Labs' FLUX 3 image family, a 5.6B-parameter model that does text-to-image, editing and multi-reference generation on your own machine.

open weights, FLUX Kommunity License: commercial use only if your annualised revenue is under US$5,000,000 · needs a single consumer GPU, down to an RTX 3090

Drama 3

A text-to-speech model you direct in plain language: alongside the script you can ask for a tone, a sound effect, or emphasis on particular words.

hosted service

Any Time, Any Space (ATAS)

A framework for making long, multi-shot videos that keeps a running record of the scene's geometry and events, so places and objects stay consistent from one shot to the next.

open source, mit · needs an NVIDIA GPU with about 20 GB free memory for the geometry service; the client side needs no GPU

2026-10-11 AI Search

WorldPlay 2

A model that generates an interactive 3D world you can move through in real time, steered either by direct controls or by typed instructions, and which stays consistent over long sessions.

open weights, CC-BY-NC-4.0 (non-commercial only); a paid hosted demo is offered separately through Reactor · needs a CUDA-capable NVIDIA GPU; ships configured for 8 GPUs, single-GPU use is supported with config changes

Terra

A research pipeline from EPFL that reads the ground surface out of human motion-capture clips and trains a simulated, muscle-powered body to walk the same terrain.

open source, apache-2.0; needs the SMPL-H body model and AMASS motion data obtained separately under their own non-commercial terms · needs an NVIDIA GPU for training and evaluating policies; loading motion data works on CPU

Room temp magnetic semiconductors

A write-up from Vals AI in which AI agents proposed two materials for next-generation computer memory: one newly designed compound that has never been made, and one existing 1999 material. Both are computer predictions only; the blog says neither the band gap nor the spin sorting has been measured in a lab.

research project; input files, raw outputs and analysis code on GitHub

OpenAI math breakthroughs

A collection of 719 mathematics write-ups produced by an unreleased OpenAI model on open research problems. Despite the "700+ breakthroughs" framing, it is a dump of manuscripts and Lean proofs rather than code or a released model, and OpenAI states only about 42% of the headline results are formally verified and that the rest "could have issues".

open source, apache-2.0

Nano Banana 2.1

Google's latest image generation and editing model, part of the Gemini 3 family; it takes text and images in and returns images up to 4K.

paid service; free to use in the Gemini app, paid through the API · needs none, runs in the cloud

MSFlow

A text-to-human-motion generator that works directly on full-resolution motion data instead of a compressed code, which lets you pin a specific joint or frame to a chosen position at generation time without retraining for it.

open source, cc0-1.0, though the authors warn the repository is only temporarily public

Kandinsky 6.0

An open family of models that generate a 5-second video together with its matching 44kHz soundtrack, including lip-sync, in one go. Comes as a 3B "Lite" and a 29B "Pro" version, plus a separate latent upscaler for up to 4x resolution.

open source, mit · needs an NVIDIA GPU and Python 3.13 or 3.14; the repo benchmarks on cards from RTX 4090 up to H100

Iris 3B

A 3-billion-parameter open text-to-image model that generates raw pixels directly at 1024x1024, skipping the VAE compression step nearly every other image generator relies on; the same backbone is also shown doing depth estimation and image restoration.

open source, apache-2.0, with weights on Hugging Face · needs an NVIDIA GPU with CUDA; roughly 12GB of weights for text-to-image, plus about 12GB each for the optional depth and upscaler parts

Haiku 5.5

Anthropic's smallest and fastest current model, aimed at high-volume cheap work like classification, summaries and support replies rather than hard reasoning — it sits well below Sonnet and Opus on intelligence rankings.

free to use in the Claude app; paid per token on the API, from $0.10 in / $0.50 out per million tokens · needs none, runs in the cloud

GPT intelligent UI

OpenAI rolled GPT-6 out to every ChatGPT tier, including free accounts, and added "Intelligent UI" — answers that can contain buttons, charts, forms and small interactive tools rather than only text. Free and Go accounts get a model called GPT-6 Luna; paid tiers get the stronger GPT-6 Sol.

free to use, with a weaker model (Luna) on the free tier and Sol on paid plans · needs none, runs in the cloud

FlashDexRetarget

A way to convert whole libraries of filmed human hand movements into working robot-hand trajectories at once, by training a single policy conditioned on the reference motion instead of optimising each clip separately — the authors claim about 100x less training compute than the previous approach.

open source, mit · needs about 110GB of GPU memory to train, or about 16GB GPU plus 100GB system RAM; also needs Isaac Sim 5.1 / Isaac Lab 2.3 and separately-registered MANO hand models

Embedding Gemma 2

A small Google model that does not write anything — it converts text, images, audio and video into comparable numeric fingerprints, so you can search a mixed pile of files by meaning. At 740M parameters it is meant to run on a phone.

open weights, apache-2.0 · needs runs on-device: about 191MB of active RAM for text only, about 567MB for the full multimodal model when quantised

DistScene

A research method that builds a 3D scene from an image by first generating the individual objects well and then distilling them into a coherent whole scene. Neither code nor checkpoints are out — both are listed as coming soon.

paper only, code and checkpoints announced as coming

CoDimRecon

A research method that turns ordinary multi-view photos of a room into a 3D scene you can actually simulate, keeping rigid objects, hinged parts like drawers, and floppy things like cables and cushions as separate kinds of geometry. Nothing is downloadable yet — the page only promises code later.

paper only, code announced as coming

Botanic 1

A pipeline for working out which mutation in a plant's DNA causes a given trait. Gemma 4 does the reasoning and planning, and a genome model called BOTANIC-1, trained on 320 plant species, scores how evolutionarily important each stretch of DNA is. Despite being showcased on a Google DeepMind page, BOTANIC-1 is built by Living Models, an outside startup — Google's part is Gemma 4.

open weights, research use only (Living Models Research Licence); the four BOTANIC-1 checkpoints are on Hugging Face · needs the demo runs locally via Ollama on a single NVIDIA L4 GPU

AlphaProtein Novo

A Google DeepMind system that designs entirely new enzymes from scratch: given a target chemical reaction it generates both the protein's 3D shape and its amino-acid sequence, then checks the result with AlphaFold 3. The repository is code only — the trained weights are a separate download under Google's own terms, not Apache.

open source, apache-2.0 for the code; model weights under Google's own terms of use · needs a CUDA 12 GPU is recommended (CPU is possible); scales across several GPUs if present

ALoDLM

Amazon's 8B text model that writes by diffusion rather than one token at a time: it drafts a whole block of tokens at once and keeps re-refining only the positions it is unsure about, so it can spend less compute on the easy words.

open weights, non-commercial only (CC BY-NC 4.0) · needs a CUDA GPU with BF16 support; the fast inference path was tested on an A100 40GB

2026-10-07 Mickmumpitz

🔗 ComfyUI Installation Guide

A free step-by-step guide on the video maker's own site for installing ComfyUI, the open-source node-based program used to run image and video models on your own PC.

free to use

2026-10-06 AI Search

universal modder

A bundle of skills, tools and shared notes that lets an AI coding agent mod PC games you own — working out the engine, writing the mod and generating art and sound; despite the repo's one-liner it supports Codex, Cursor, Gemini CLI, Copilot and OpenCode as well as Claude Code.

open source, mit

rea

A toolkit that hooks a coding agent up to reverse-engineering tools so it can work out how an app's feature is built without its source code; it needs Node 22.19+ and an external disassembler, either the free Ghidra or the commercial Hopper.

open source, mit

PortPS5

An early tool that turns PS5 game dumps you have already decrypted yourself into Windows programs, running the game's code on your own CPU and translating its graphics to Vulkan; it is a fork of AnyPS5 and is pre-1.0, with its five target games not yet playable start-to-finish.

open source, gpl-2.0 · needs Windows 10 or 11 with a Vulkan 1.3 GPU

ilspy

A long-established free tool that turns compiled .NET programs back into readable C# source.

open source, mit

ida mcp

Hex-Rays' own connector that lets an AI agent drive IDA, the commercial disassembler; the connector is free but you need a paid IDA 9.4 or later licence to use it.

open source, mit

ghidra mcp

A plugin that exposes over 200 tools from Ghidra, the free NSA-built reverse-engineering suite, so an AI agent can analyse binaries through it.

open source, apache-2.0

Cpp2IL

A tool for recovering readable code from Unity games compiled with IL2CPP; its authors describe it as still a work in progress.

open source, mit

AnyPS5

The original project for converting decrypted PS5 game executables into native Linux and Windows programs without emulation; only a short list of games is verified working, with a 2D platformer as the headline result.

open source, gpl-2.0

aisearch oct

Not a launch: the link is the video's sponsor tracking URL and redirects to Runway's ordinary homepage, which sells cloud tools for generating and editing video, images and audio, plus a developer API and a robotics simulation toolkit.

paid service, free tier of 125 one-time credits

ai game modding guides

Sixteen written guides on getting AI coding agents to build game mods — choosing an agent, prompting, testing, legal risk — with templates and worked examples; it is documentation only, no tool, and the repo labels itself a draft.

open source, MIT

2026-10-04 AI Search

Whistle

A 16.9 MB speech-to-text model that runs on an ordinary CPU with no GPU, small enough for phones, wearables and microcontrollers, covering seven languages.

open weights, Apache-2.0 · needs CPU only, no GPU needed; 16.9 MB model file

Tactile Step

Puts pressure sensors in a humanoid robot's soles so it lands its feet more softly and stays stable on uneven ground; shown on a Unitree G1.

research project

Sol Refiner

NVIDIA's one-step upscaler that takes a low-resolution video draft from another generator and finishes it at high resolution, reported up to 27x faster than generating at full size.

open source, Apache-2.0; code only, no weights published

Robot badminton

Teaches a humanoid robot to rally at badminton, including jumping returns, learned from a small amount of human motion capture.

paper only

Point2Part

Cuts a 3D object into parts from one clicked point per part, with the pieces guaranteed not to overlap or leave gaps; works from a mesh or a single photo.

research project, code on GitHub with no licence stated

Plasma conjecture

A Lean 4 proof script machine-checking a new result that disproves Grad's 1967 conjecture about symmetry in stellarator fusion plasmas; the video said GPT-6 Astra found the counterexamples, but neither the repository nor the paper's abstract credits any AI.

research project

PixelUMM

NVIDIA's single model that both reads and generates images and video as raw pixels, with no separate image encoder or VAE, on a Qwen3-8B backbone.

open weights, NVIDIA non-commercial licence; code Apache-2.0

Phonon 2

A 164 MB speech-to-text model that transcribes about an hour of audio in twenty seconds on a MacBook Air.

open weights, CC-BY-4.0 · needs runs on Mac, Linux, Windows or NVIDIA GPUs; 164 MB

PDMD

A ByteDance and UC San Diego trick that gets a video model down to four generation steps by projecting the critic's own error out of the training signal, released as a distilled MiniMax-H3.

open weights, Apache-2.0

PAMI

Generates a person and the object they are handling moving together from a text prompt, tying the object to wrists and ankles so it stops floating away or passing through the body.

open source, MIT

OpenDots

A self-hostable template from CopilotKit for always-on agents that work across chat, phone calls and Slack; despite the name it does not describe itself as a clone of OpenAI's Dots.

open source, MIT · needs Node.js 24 and npm, plus your own model provider API key

Open Dots

A self-hosted stand-in for OpenAI's Dots with chat, connectors, approvals and optional computer use, which its own README calls a prototype that is not feature-equivalent and not ready for multi-user hosting.

open source, MIT · needs Node.js and npm, Python 3.10+, your own inference API key; Docker for the computer-use runtime

Olmo Core 3

Ai2's open training stack, rebuilt to train mixture-of-experts language models up to trillion-parameter scale about 2.7x faster than its previous version. It is plumbing for people who train models, not a model you can use.

open source, Apache-2.0

Napoleon code cracked

A hobbyist's writeup of using GPT-6 Astra to transcribe and crack the cipher in an 1809 letter to General Marmont, then redate it to late March 1809; the author says he has never broken a cipher before, and calls it one reading of one letter from one low-resolution plate, with five symbols still weakly evidenced.

free to use

IQuest Q1

A large mixture-of-experts language model aimed at agentic coding and multi-step tool use, with 320 billion parameters in total and 15 billion active per token. Weights are published on Hugging Face.

open weights, modified MIT: commercial products must display "IQuest-Q1" in their interface · needs the published serving example runs across 8 GPUs (vLLM tensor-parallel-size 8)

Inspatio World 1.5

A model that takes a photo, a video or a 360-degree panorama and generates new camera views of the same scene, so you can move around in it. Code and weights are released and there is a live demo.

open weights, Apache-2.0

Ideogram 4.5

Ideogram's image editing model, which keeps the rest of a picture intact over repeated edits and can change detail inside very large images without shrinking them first. Available through its website, iPhone app and API.

hosted service

GPT 6.1 Sol

OpenAI's cheaper near-frontier model for coding, computer use and document work, priced at $2 per million input and $10 per million output tokens.

paid service · needs none, runs in the cloud

Gemini 4 Argon

Google's new top-end Gemini model for coding, office knowledge work and cyber defence, with a one-million-token context. Despite the launch framing, you cannot get it yet: access is limited to vetted cyber-defence testers in Google's Fairwind programme, with paid API and AI Ultra subscribers to follow.

paid service, limited early access only · needs none, runs in the cloud

Flux 3

Black Forest Labs' image generator and editor, whose selling point is that you place each element in a box on the canvas and edit a finished picture one box at a time. There is a paid API and a free web playground.

open weights, non-commercial only; a separate paid licence covers commercial self-hosting

Elevenlabs V4

ElevenLabs' latest text-to-speech voices, in two versions: v4 for produced audio with multiple speakers, and v4 Turbo for live use at around 100 ms latency.

hosted service, free tier of 10,000 credits a month; paid plans from $6/month · needs none, runs in the cloud

Dots

Always-on ChatGPT agents you set up once with a goal instead of starting a fresh chat, each with its own cloud computer; announced at DevDay on 29 September 2026, and not offered to Pro users in the EEA, Switzerland or the UK.

paid service · needs none, runs in the cloud

DMAD

A training trick that compresses an image or video generator from the usual 50-plus denoising steps down to one to four, so it produces results far faster at similar quality. Code and weights are published.

open source, Apache-2.0

Comfy Agent

An assistant built into ComfyUI that takes a plain description and wires up, runs and tweaks the image or video workflow for you instead of you connecting the nodes by hand. It runs on Comfy Cloud today; the local desktop version is promised in a few weeks.

hosted service, free to try then spends Comfy Credits

Claude Sonnet 5.5

Anthropic's mid-priced assistant model, a faster and cheaper replacement for Sonnet 5 at $2 per million input tokens and $10 per million output tokens. Also available in the free tier of the Claude app.

paid service · needs none, runs in the cloud

AstaBrief

An 8-billion-parameter model from Ai2 that turns a research question and a stack of papers into a written report with citations; it powers the fast mode of Ai2's Asta tool and the weights can be downloaded and run in-house.

open weights, Apache-2.0

Arex 2

A 27-billion-parameter agent model from BAAI that improves its own answer over several rounds, proposing a solution, measuring it, then revising. Aimed at coding, machine-learning engineering and research tasks rather than the image work its Hugging Face tag suggests.

open weights, Apache-2.0

analytic 3d equilibria

Scripts and data backing Matt Landreman's plasma-physics paper on exact 3D magnetic-equilibrium solutions that are counterexamples to Grad's conjecture; the repository includes the GPT-6 Astra Pro chat in which the solutions were found, though the arXiv paper itself does not mention AI.

open source, mit

2026-09-27 AI Search

WorldCrafter

Tencent ARC's world model: give it one image or a text prompt and you can fly a camera through the 3D scene it invents, and it stays consistent when you look away and back. Code and weights are public, but Tencent's licence allows academic use only, not commercial or production use.

open weights, academic use only · needs an NVIDIA GPU; no VRAM figure is given

Supra2 Img

A deliberately tiny text-to-image model - about 105 million parameters, trained from scratch on synthetic images - that produces 256x256 pictures. The model card makes no claim about running it on a phone.

open weights, Apache 2.0

Step 5

StepFun's flagship cloud model for agent and coding work, taking text, images and video with a one-million-token context. It is an API only - there are no weights to download.

paid service · needs none, runs in the cloud

SkildAI soccer

A research write-up, not a release: Skild had its S1 robot model learn to play soccer by competing against copies of itself in simulation, then moved the resulting policy onto a real humanoid. No code, weights or product are offered.

research project

Project Suncatcher

Google's early-stage study of whether AI chips could run in orbit on solar power. A prototype satellite carrying TPUs has launched; a two-satellite laser-link test is planned for 2027. There is nothing here to use.

research project

Ovis Embedding

Three embedding models (2B, 3B and 9B) that map text, images, video and scanned documents into one shared vector space, so you can search across media types with a single query.

open weights, Apache 2.0

OpenMuse

Self-hosted software for a long-running personal agent with its own browser, terminal and files. It is a harness, not a model: you supply your own OpenAI, Anthropic or Google API key, plus a CopilotKit Intelligence project key, so running it is not free.

open source, MIT · needs Node 24 LTS and pnpm; Docker as well if you want the Linux terminal

Mira Scene

Turns a single photograph into an editable 3D scene - separate object meshes, their layout, and a matching environment map. The linked Eden page does not load; the code and a pretrained checkpoint are published by VAST AI Research, which states no licence for them.

open weights

Ming Image

A 6-billion-parameter text-to-image model built for text-heavy design work - interfaces, posters, infographics - at up to 2K resolution, and it can output transparent backgrounds.

open weights, MIT · needs one CUDA GPU with 80 GiB of VRAM is the configuration the authors validated

Mimo V2.6

Xiaomi's newest model family - a Pro and a Flash model plus a distilled 9B - handling text, image, video and audio with up to a million tokens of context. Xiaomi also published over 7,000 reinforcement-learning environments and its training framework alongside the weights.

open weights, MIT

Limite 1B

A 1-billion-parameter model built for one job only: solving hard competition maths problems in one shot, at high throughput. It is not a chat assistant.

open weights, Apache 2.0 · needs Linux x86-64 with an NVIDIA GPU and a CUDA 13.0-compatible driver

Grok 4.7

The newest Grok model, pitched at coding and knowledge work. The API is listed at $2 per million input tokens and $6 per million output tokens, with a faster variant at double that; there is also a free tier through Grok Build. No weights are released.

paid service · needs none, runs in the cloud

GPT 6 Sol Luna

OpenAI's cheaper additions to the GPT-6 line: Sol is the middle tier aimed at coding and agent work, Luna the cheapest for high-volume simple jobs, at roughly half the API price of their GPT-5.6 equivalents. Paid ChatGPT plans get both in ChatGPT Work and Codex; free and Go users get Luna in the desktop app.

paid service · needs none, runs in the cloud

Gene discovery

A research write-up, not a product: Anthropic's life-sciences group had Claude agents comb DNA databases and surface a previously undescribed family of 'array-associated reverse transcriptases' that look CRISPR-like. Anthropic says plainly that it does not yet know what the system does, and that humans did all the lab work. A technical preprint is out; there is nothing to use.

news article

Gemini 3.8 TTS

Two Google text-to-speech models that build a voice from a written description, take line-by-line direction on how each line should be delivered, and can copy a voice from a 30-second sample, across more than 100 languages.

paid service · needs none, runs in the cloud

GAE

A research system that generates a 3D-consistent world: alongside the video it produces a depth map, the camera path and a point cloud of the scene, so things stay where they were when the camera looks back. Code and the 1B weights are out, but Tencent's licence restricts use to academic purposes only.

open weights, academic use only

Flux 3 Action

Black Forest Labs' first robot model rather than an image one: a 7B model that takes the robot's camera views, its joint positions and a written goal, then predicts both the next batch of motor commands and what the scene will look like once they are carried out.

open weights under the FLUX Kommunity License v1.0, which permits commercial use only for organisations under $5m annual revenue

Dots3 Note

The first open model in Red Note's dots3 line: a 280B mixture-of-experts with 16B active per token that takes text, images, video and audio in and writes text out, with a 512K-token context window. The video understates it as 28 billion parameters and a 512-token context.

open weights, Apache-2.0 · needs an 8-GPU node for the FP8 version; BF16 needs more memory

CLM

An 8B model that scores a menu of candidate next actions and picks the one that fits the situation, instead of writing text. It is two small trained heads on top of a frozen Qwen3-8B, which is how it answers up to nine times faster than the models it is compared against on computer-use and tool-calling tasks.

open weights, Apache-2.0

Aikido Altar

A shrunken copy of GLM-5.3, 504B parameters with about 40B active, tuned for security work so a company can run vulnerability-finding and pentesting on its own hardware instead of sending its code to an outside service. The compression from 1.51TB down to 328GB costs some accuracy: 60.4% recall on Aikido's own CVE benchmark against 65.6% for the uncompressed original.

open weights, inherits the GLM-5.3 licence · needs a 4x H200 node; 328GB of weights

2026-09-24 AI Search

Claude Opus 5.5 release page

Anthropic's updated top-of-the-range assistant model, aimed at coding and long unattended tasks. Per-token prices are 20% below Opus 5; the "40% cheaper" figure is Anthropic's own test result at default settings, not the list price.

paid service · needs none, runs in the cloud

2026-09-22 AI Search

Workflows and models

Alibaba's Qwen-Image 2.1 image generator and editor, repackaged so it drops straight into ComfyUI — model weights, text encoder, VAE and ready-made workflows. The video calls it open source, but the weights carry Qwen's research licence rather than an open-source one.

open weights, Qwen Research License

GGUFs

Community-made compressed (GGUF) copies of Qwen-Image 2.1, so the same image generator runs on an ordinary consumer graphics card. ComfyUI workflows are included.

open weights, Qwen Research License · needs 6–8 GB VRAM for the 4-bit versions, 12 GB+ for the near-lossless 8-bit one

Anime consistency lora

One experimental add-on for Qwen-Image 2.1 that helps keep an anime character looking like themselves across edits. It is not an image generator in its own right — it needs the main model.

open weights, apache-2.0

2026-09-20 AI Search

ZGCM

A 7.39-billion-parameter language model built for maths problems and tool-assisted search, released with its training code as well as its weights.

open source, MIT

ZAI infra

Z.ai's engineering write-up on serving its GLM-5.3-Flash model from scratch on more than 100,000 Chinese-made accelerators, in which an agent driven by GLM-5.3 did much of the kernel optimisation itself.

news article

Xiaomi realtime RL

A live public dashboard from Xiaomi showing its mimo-v2.6-pro and mimo-v2.6-flash models being trained by reinforcement learning as it happens — reward curves, step timings and running compute cost. It is a window on a training run, not a model you can download.

free to use

R2T2

NetEase Youdao's 2-billion-parameter speech-to-text model (Confucius4-R2T2) that transcribes as you speak and never rewrites words it has already committed. The site calls it open source, but only the code is - the weights ship under NetEase's own licence.

open source, Apache-2.0 code; the model weights are separately under the NetEase Model Use License Agreement

Qwen 3.8 Omni Flash

Alibaba's multimodal Qwen that takes text, images, audio and video and replies in text, with a one-million-token context and much cheaper audio input. Unlike most Qwen releases it is API-only - no weights were published at launch, despite some coverage saying otherwise.

paid service · needs none, runs in the cloud

Qwen 3.8 Live Translate

Alibaba's live interpreting model: it translates a speaker into another language while they are still talking, with about 2.3 seconds of lag, across 60 input languages, and can tell separate speakers apart.

paid service, sold as a streaming API; the weights are not published · needs none, runs in the cloud

OpenAI hacked

A security firm's write-up of how it chained a bug in an image-handling library with a single sign-on misconfiguration to reach OpenAI employee accounts and internal repositories. OpenAI patched it in about 14 hours and paid a $6,500 bounty.

news article

Odyssey 3

An announcement of a single model trained on video that is meant to control robots, cars, drones and game characters alike. Worth noting against the launch claims: nobody outside the company can try it yet — the page promises a public release 'in the coming weeks' and names no licence, price or format.

unavailable

Occamy

A downloadable model aimed at long multi-step office and tool-use work rather than quick answers; it can read images as well as text. 35 billion parameters in total but only about 3 billion active at a time, so it is cheaper to run than its size suggests.

open weights, Apache 2.0

Nimble

Not a model in its own right despite the name: a 165 MiB adapter that sits on top of Qwen3.5-9B, which you download separately. It makes the model pick from a fixed list of answers or score against a rubric instead of writing prose.

open weights, Apache 2.0 · needs a CUDA GPU with BF16 support

Minimax code

A coding assistant that runs in your terminal, reading a project and editing files on request. The tool itself is free, but it needs a model behind it — either MiniMax account credits or your own API key from another provider.

open source, MIT

Meridian

A video model that re-shoots an existing clip or still from a new camera angle or at a different speed. Two points the licence makes that a demo reel would not: it may not be used in the EU, the UK, South Korea or the United States, and its geometry component is non-commercial research only.

open weights, non-commercial geometry component and barred in the EU, UK, South Korea and the US · needs one B200 GPU; 88-113 GiB of memory in the published runs

Laya

A small, fast classifier that sorts text into categories and returns a confidence score — routing a support ticket, flagging a phishing email — in about 33 milliseconds. It does not write text, so it is not a chatbot.

open source, Apache 2.0 · needs a single consumer GPU; CPU inference also supported

Jing Dao

A research preview of two pieces: JING, which generates first-person video and sound for a world you walk around with the keyboard, and DAO, the shared-world engine meant to sit behind it. Only JING has been released; DAO is described, not shipped.

open weights, MiniMax H3 Community License (JING only) · needs the released demo was validated on six H100 GPUs

Jev

A model that returns typed, structured values with confidence scores instead of free text, meant for making decisions inside software rather than for chatting.

paid service, early access by waitlist · needs none, runs in the cloud

In context learning

A research method for making a general vision-language model drive a real robot arm: the model is shown a human video, a demonstration or a goal picture as context, and a small controller turns its replies into arm movements. Nothing is retrained.

research project, licence deliberately withheld — the repository states redistribution and commercial use are not granted

Gemini 3.8 Live

Two Google voice models for real-time spoken conversation: a cheaper one that listens, sees and detects the language being spoken, and a slower variant that reasons through multi-step problems mid-conversation. No weights are published.

hosted service · needs none, runs on Google's servers

Dream RSI

A research paper on letting an agent replay its own past experiments as a cheap offline simulator, so it can test thousands of improvements to itself without re-running the expensive real thing. Despite the working demo on the page, the code is not out: the GitHub repository says only that a release is being prepared.

paper only

Decrypting German enigma

A write-up of a two-day effort that recovered the settings and plaintext of a genuine 82-letter German Enigma message from July 1941, with an interactive Enigma simulator and downloadable code and search records so anyone can check the result.

free to use

Bonsai 2 27B

A squeezed-down copy of Qwen3.8 27B that stores each weight as roughly 1.76 bits, so the whole model is a 5.9GB download while keeping about 98% of the original's benchmark scores.

open weights, Apache 2.0 · needs 5.9GB of memory; runs on NVIDIA GPUs via CUDA or Apple silicon via MLX

Astra for Law

OpenAI's GPT-6 Astra model preset for legal work, with a search index over US case law, statutes, regulations and court rules, and plugins from legal software vendors.

paid service, no public price; access is limited to API customers and to selected large firms through OpenAI's Trusted Access programme · needs none, runs in the cloud

aisearch sept

Not a new release — this is the video's sponsor link, which redirects to Runway's ordinary homepage. Runway sells a cloud suite for generating and editing video, images and audio.

paid service, free tier available

2026-09-18 AI Search

Deepseek V4.1 Flash tech report

DeepSeek's technical paper on DeepSeek-V4.1-Flash — not the model itself, but the write-up of how it was built: a 552-billion-parameter mixture-of-experts model that handles images as well as text, yet only uses 8 billion parameters to read your input and 16 billion to write the answer, with a new attention scheme that cuts the memory it holds per token of context to roughly 890 bytes, about a quarter of the previous DeepSeek generation.

open weights, MIT; the paper sits in the same Hugging Face repo as the downloadable weights

2026-09-15 AI Search

YuE2 official repo

A music generator that turns a style description and lyrics into a finished song, writing an editable score first and then performing it, and that can also transcribe an existing recording and redo it in another style. The project's own requirement is a 24 GB NVIDIA card; the video's "4 GB of VRAM or less" applies to a shrunken community repack, not to this code.

open source, apache-2.0; model weights separately cc-by-nc-4.0, non-commercial only · needs Linux, Python 3.12, NVIDIA GPU with BF16 support and 24 GB VRAM

YuE2 models

The YuE2 music model and its audio encoder repackaged as single files for ComfyUI: a 7.8 GB full-precision checkpoint, a 3.96 GB shrunken one, and a 1.4 GB encoder. The page states no memory requirement at all, so the "under 4 GB" figure quoted in the video is the size of the smaller download rather than the graphics memory needed to run it.

open weights, non-commercial only (cc-by-nc-4.0)

Comfy workflow

The ready-made ComfyUI setup file for YuE2 — a saved wiring diagram connecting the model, encoder and sampler, plus a second branch that converts an uploaded song into notation for the cover-version feature. It is a plain attachment download with no licence stated on it.

free to use

2026-09-13 AI Search

YuE2

A music generator that first writes a score and then performs it with vocals and backing, which the authors say matches Suno v5/v6; the weights are non-commercial only.

open weights, cc-by-nc-4.0 non-commercial only; code apache-2.0 · needs NVIDIA GPU with 24 GB VRAM and BF16 support

World Sculpt

Splits a cluttered photographed or generated scene into hundreds of separate, complete 3D object meshes you can move one at a time; it needs per-object masks and 3D boxes as input, not just a bare photo.

open source, apache-2.0

Unimate

Generates animation for any rigged 3D character from a text prompt, including non-human skeletons, without retraining per skeleton. The code and the 13,000-clip training set are out, but the pretrained model itself is still listed as coming soon, so you cannot run it as shown yet.

open source, mit

UnifoLM

Unitree's 6-billion-parameter foundation model for humanoid robots, trained on about 2,500 hours of real-robot recordings so one model can run 64 different tabletop and whole-body tasks. The headline WLA model is not actually downloadable yet: so far only two smaller reasoning models (UnifoLM-ER-1 and ER-Flow) and some datasets have been released, and the GitHub repo holds documentation rather than code.

open weights, apache-2.0 (the two released models only)

UMR

Converts human motion capture into trajectories a humanoid robot can follow, learning the body-to-robot correspondence from point clouds instead of hand-written joint mappings. Code is published but no trained weights and no licence are given.

open source, no licence stated

Suno V6

Suno's new music generator, built with Warner Music Group, BMG and Believe, which also brought in screening of uploaded audio and lyrics for unauthorised use.

paid service; v6 and v6-wild need a Pro or Premier subscription, a smaller v6-mini is free to everyone · needs none, runs in the cloud

Show Harness

A thin control layer that lets an off-the-shelf vision-language model drive a robot by picking from a small set of named actions, instead of training a dedicated robot policy.

open source, code and adapters on GitHub and Hugging Face, apache-2.0; the Gemma-based variant falls under Google's Gemma terms instead

RealSWE

A leaderboard scoring AI coding agents on tasks drawn from private company codebases. You can read the scores, but the tasks and the code behind them are not public and sample tasks are only available on request.

unavailable, leaderboard only

Navier Stokes

OpenAI says a swarm of its agents produced, and checked in the Lean proof language, a proof that 3D fluid equations can blow up in finite time. It is an announcement, not a settled result: the Clay Institute still lists the Millennium Prize problem as unsolved and the work has not been independently peer reviewed.

news article

MiniCPM5 2B

A 2-billion-parameter language model built to run locally on phones, laptops and other constrained devices, with a 128k-token context and ready-made quantised builds.

open weights, apache-2.0

Marigold v2

Estimates depth, surface normals and other per-pixel properties from a single photo by cheaply fine-tuning an existing image-editing diffusion model. This link is the project's page on Hugging Face; the code and the LoRA weights live in its own repositories.

open weights, apache-2.0

Lingbot World 2

A world model from Ant Group's robotics arm that generates an explorable 3D-looking environment on the fly at 720p and 60fps, staying coherent as you keep moving through it.

open weights, cc-by-nc-sa-4.0 non-commercial only, so despite the open-source framing you cannot build a product on it

Isaac 0.5

A 36B sparse mixture-of-experts model for robots that does video understanding, spatial reasoning and the actual motor control in one network, trained across 35 different robot bodies.

open weights, apache-2.0

GPT for finance

A version of ChatGPT Work aimed at investment bankers and equity researchers, with paid market-data feeds (Daloopa, PitchBook, LSEG News, Crunchbase) wired in so figures in its answers can be traced back to a source.

paid service; sold to firms through OpenAI's sales team, no public price · needs none, runs in the cloud

Fire3D

Turns a single photo or a casual video clip into separate, complete, textured 3D objects you can move individually, in under a minute; unlike most such tools it needs no masks, boxes or prompts to pick the objects out.

open source, code MIT; the released checkpoint bundle is tagged 'other' because it also carries DINOv3 and TRELLIS.2 files under their own upstream licences · needs the authors' timings are on one 80 GB A100

Edge0

Runs big mixture-of-experts language models on a Mac by streaming in only the experts each token needs, so a 35B model works within roughly 3 GB of active memory.

open source, apache-2.0 · needs macOS on Apple Silicon (M1-M4); about 23 GB of disk for the 35B model, 4.2 GB for the 8B

Deepseek v4.1 Flash

DeepSeek's small, cheap model in its new causal encoder-decoder family: 552B total parameters but only 8B active reading your input and 16B writing the answer, and it reads images as well as text.

open weights, MIT, on Hugging Face; also sold through DeepSeek's own API, where off-peak rates are half the peak ones

AuK

Tencent's 1.5-billion-parameter speech model that generates, rewrites, cleans up and separates speech from plain-language instructions, with a faster four-step variant.

open source, mit · needs about 25 GB of GPU memory, or about 17 GB with CPU offloading

AlphaGenome Atlas

A Google DeepMind database of predicted biological effects for all 9 billion possible single-letter changes in human DNA, browsable through a web portal with no coding; the announcement does not say what it costs or who may use it.

hosted service

2026-09-11 Mickmumpitz

Wan 2.2

Alibaba's open family of video-generation models — text, image or audio in, video out — and the engine 4DAnyone is built on top of.

open source, Apache-2.0 · needs the 5B model needs 24 GB of VRAM (an RTX 4090 will do); the 14B models want 80 GB, less with the offload flags

The PAPER

The code for 4DAnyone, which takes one handheld video of a person and invents the dozens of synchronised camera angles a studio rig would have shot, then turns them into a 4D Gaussian splat you can fly a virtual camera through.

open source, Apache-2.0 · needs under 24 GB of VRAM at peak; about 27 seconds for a 121-frame clip on one RTX 4090

Nerfstudio

A free toolkit from Berkeley researchers for training, viewing and exporting NeRFs, offered here as the alternative place to train the splat.

open source, Apache-2.0 · needs an NVIDIA card with CUDA installed; tested against CUDA 11.8

gsplat

The CUDA library that does the actual rendering maths for Gaussian splatting, used here as the splat trainer.

open source, Apache-2.0 · needs an NVIDIA card with CUDA, and PyTorch 2.7 or newer

4DAnyone (Paper)

The research paper behind 4DAnyone, from Ant Research and Zhejiang University, explaining how it keeps dozens of generated camera angles consistent with each other.

paper only

4DAnyone (Model)

The downloadable weights for 4DAnyone. They are a bundle under several different licences rather than one: the main checkpoint is Apache-2.0, but the Wan 2.2 speed-up LoRA shipped alongside it is non-commercial and the GVHMR pose checkpoint is research-only — so the commercial catch the video pins on GVHMR alone also applies to the speed-up LoRA it recommends.

open weights, mixed licences — the main checkpoint is Apache-2.0, but parts of the bundle are non-commercial or research-only

3DGS Render by KIRI Engine

A free Blender add-on for importing, editing, animating and rendering Gaussian splats, now maintained by the community rather than KIRI Engine.

open source, GPL-2.0

2026-09-10 AI Search

GPT Image 2.5 release page

OpenAI's newest image generator and editor, in ChatGPT on every tier including the free one, and in the API as two versions - Flare for fast everyday images and Sunburst for precise editing - which OpenAI's own docs list at identical token prices.

hosted service

2026-09-07 AI Search

GPT 6 release page

OpenAI's new flagship model, pitched at computer use, coding and science, rolling out to paid ChatGPT tiers and the API.

paid service · needs none, runs in the cloud

2026-09-06 AI Search

WeatherNext 3

Google's global weather forecasting model, now behind the forecasts in Search, Gemini and Maps; developers can pull the forecast data from BigQuery, Earth Engine or Cloud Storage.

hosted service · needs none, runs in the cloud

Viggle Animate

Replaces the character in an existing video with a different one, keeping the original motion and timing; it is an add-on to MiniMax-H3 rather than a standalone model.

open weights, MiniMax-H3 community licence · needs more than 80GB of VRAM at bf16 (a 96GB card, or offload blocks to system RAM and run slower)

VideoDeltaNet

A reworked attention design for the MiniMax-H3 video model that makes long clips much faster to generate; code and weights are published.

open weights, MiniMax-H3 community licence, which excludes the US, EU, UK and South Korea · needs the quoted speeds use 8 NVIDIA B200 GPUs

TimesFM

A small (330M-parameter) Google model that forecasts several related time series at once without being trained on your data; the weights are downloadable but the licence forbids commercial use.

open weights, non-commercial only

SolarWM

A video world model trained on an openly released 1.4M-clip dataset, which extends five-second clips into much longer interactive simulations; code, data and several checkpoints are all published.

open weights

Qwen 3.8 0902

A dated snapshot of Alibaba's largest Qwen model, sold through its cloud API at $2 per million input tokens and $6 per million output.

paid service · needs none, runs in the cloud

Lucida

Turns photos of a real room into separate, editable 3D objects, filling in the parts the camera could not see; the page is a draft with no code or weights released.

research project

LLaDA image

A 6B-parameter open model that both generates images from text and edits existing ones from written instructions.

open weights, Apache 2.0

Intern Lumina U2

A single model meant to handle image and video understanding, image generation and image editing together; only the inference code is out so far, and the Hugging Face repository has no weights in it yet.

research project

H3 World

Turns the MiniMax-H3 video model into something you steer with keyboard keys, moving a character and the camera through a generated scene; it is a small add-on layer, so you also need the 33B base model.

open weights, Apache 2.0, but requires the MiniMax-H3 base weights

GWM 2

Runway's system for generating an interactive world with sound as you move through it; the page calls it a research preview and admits real-time speed costs image quality, with no public sign-up.

research project

GPT 6 Astra

OpenAI's new flagship model, pitched at computer use, coding and science, rolling out to paid ChatGPT tiers and the API.

paid service · needs none, runs in the cloud

Gemini 3.8 Flash

Google's fast, cheap Gemini tier, updated for coding and agent work, plus a security-focused 'Cyber' variant. The video says both are already out via the API; Google's post restricts the Cyber model to vetted defenders who apply to its Fairwind programme.

paid service · needs none, runs in the cloud

Fly brain

A complete wiring map of a male fruit fly's brain and nerve cord — 166,000 neurons and about 125 million connections — reconstructed from electron microscope slices by AI. The paper is in Cell and the dataset is public via HHMI Janelia.

research project

Deepseek v4 flash vision

DeepSeek's first experimental multimodal model in the V4 family, a 305B-parameter system that reads images as well as text and is built for agent and tool use.

open weights, MIT · needs large: the model card's own serving example runs it across a four-GPU GB300 node

ComfyUI VDN H3

A ComfyUI plug-in for running the faster VDN-H3 version of MiniMax's H3 video model, which mixes short-range and long-range attention so cost grows linearly with clip length. The video calls it a decent speed-up for consumer GPUs; the author says outright not to use it if your VRAM or RAM is limited.

open source, Apache-2.0 for the node, but the model checkpoints carry MiniMax's H3 Community Licence, which excludes some territories · needs a single NVIDIA GPU on Windows or Linux; the node adds roughly 4.3 GB of VRAM on top of the base model, and was tested on an RTX 5090

Atlas

World Labs' world model: from text, images, video or 3D input it builds a 3D scene you can move a camera through, and can export point clouds or Gaussian splats rather than just flat video.

unavailable, early access with selected partners by request

2026-09-03 AI Search

Claude Fable and Mythos 5.1

Anthropic's newest pair of general-purpose assistant models, aimed at coding and knowledge work.

paid service · needs none, runs in the cloud

2026-08-30 AI Search

VoiceMem

Builds a memory of a spoken conversation — who was mentioned, what was said, and the speaker's preferences — from the audio itself.

open source

A robot control model that learns a new task from a single video demonstration, with no retraining.

hosted service

Qwen 3.8 Flash Next

Alibaba's fast, low-cost tier of its Qwen assistant, announced alongside its chat and image tools.

hosted service

Orbit

A benchmark for testing how well camera-position reconstruction works on real 360-degree video.

research project

One Video One World

Rebuilds a moving 3D scene from one ordinary video, with separate objects you can simulate, and no training step.

research project

Omni 1.1 Flash

Google's video generation model for developers, with added controls over how a shot is composed.

paid service · needs none, runs in the cloud

Minimax FastH3

A faster rebuild of MiniMax H3 that generates video and its audio together.

open source

Hy4

A preview of Tencent's Hunyuan 4 text model, a 770B mixture-of-experts with 49B parameters active per token and a one-million-token context window.

open weights, Apache-2.0

H3 Max

A speed-tuned version of MiniMax H3, run on fal's own servers rather than your machine.

paid service · needs none, runs in the cloud

Google PPE

Google's system for automatically building global environmental prediction models from Earth observation data.

research project

GLM 5.3 Flash

The announcement page for the faster, cheaper tier of Zhipu's GLM 5.3. The page returned no readable text to us.

unclear

GLM 5.3

Z.ai's 753B mixture-of-experts text model, aimed at coding and long-running agent work; it reuses the GLM-5.2 base model and gets its gains from further post-training.

open weights under Z.ai's own GLM-5.3 licence, which does allow commercial use; only providers with over $10bn revenue reselling model access need a review first

Gemini 3.5 Transcribe

Google's speech-to-text service, updated to handle messier audio and context.

paid service · needs none, runs in the cloud

FixAnything

Cleans up rendering errors in 3D scenes by passing them through a video model.

research project

Fibo 1.5

A text-to-image generator you prompt with structured JSON rather than a sentence, so lighting, camera and other details are set as separate fields; this version is a distilled build that makes an image in 4-6 steps.

open weights, non-commercial only

DiffusionOPSD

A training technique for image models that makes their fine-tuning cheaper to run and easier to inspect.

research project

Code World Model

A coding agent that keeps an internal model of the system it is working on.

research project

Block 3D

Generates 3D objects from a text prompt, building them block by block to save compute.

research project

2026-08-25 Mickmumpitz

Wan 2.1

Alibaba's open family of video generation models covering text-to-video, image-to-video, editing and video-to-audio, with a small 1.3B version meant for consumer graphics cards. It is the base the Mickmumpitz workflow builds on; the repo's own changelog ends at Wan2.1 VACE in May 2025, so it is not the newest Wan release.

open source, Apache 2.0 · needs the 1.3B text-to-video model needs 8.19 GB of VRAM; about 4 minutes for a 5-second 480p clip on an RTX 4090

UniSHARP Project Page

Insta360's research on turning a single photograph into a sharp, viewable 3D scene.

research project

UniSHARP (Insta360)

Insta360 research code that turns a single photo into a 3D scene you can look at from new angles, working from perspective, wide-angle, fisheye or 360-degree cameras.

research project, CC BY-NC 4.0 - non-commercial use only

Triton

A language and compiler for writing fast custom GPU code in something close to Python instead of CUDA; in these video workflows it is the speed-up people install alongside SageAttention rather than something you use directly.

open source, MIT · needs NVIDIA GPUs of compute capability 8.0 or newer, or AMD GPUs on ROCm 6.2 and later

SeedVR2

ByteDance's video restoration and upscaling models; the video uses SeedVR2 to sharpen generated frames, but the repo aims it at cleaning up degraded or AI-generated video rather than at consumer upscaling.

open source, Apache-2.0 · needs one H100 80GB for 720p video; four H100 80GB cards for 1080p and 2K

SageAttention

A drop-in replacement for the attention step inside image, video and language models that runs it in lower precision, which the authors measure at 2-5x faster than FlashAttention with no loss on end-to-end scores.

open source, Apache-2.0 · needs NVIDIA Ampere, Ada, Hopper or Blackwell GPUs, with the CUDA version matched to the card (12.0+ for Ampere, 12.8+ for Blackwell)

Postshot (Jawset)

Postshot, a Windows desktop application for building photorealistic 3D reconstructions from footage.

unclear, the site carries a pricing page · needs Windows 10 or later, Nvidia RTX 2060 or better

Poly Haven (free 360 images)

A free library of 360-degree environment photographs, usable for any purpose without an account.

free to use

NVIDIA Lyra

NVIDIA's research code for turning a single image or video into an explorable 3D world. The video's complaint that it needs 91GB of checkpoints and Linux is not something the repository itself states.

open source, Apache-2.0 code; the model weights carry their own separate licences

MoGe (Microsoft)

Microsoft's model for estimating the 3D shape of a scene, plus depth, surface angles and camera field of view, from one ordinary photograph.

open source, MIT for the code; no licence is stated for the pretrained weights · needs not supported on macOS; roughly 60ms per image on an A100 or RTX 3090

Matrix-3D Project Page

Turns one 360-degree photograph into a 3D scene you can move around in. Project page.

open source

Matrix-3D Paper

The research paper behind Matrix-3D, on generating explorable 3D worlds from a single panorama.

paper only

✓Matrix-3D (Skywork AI)

Turns a text prompt or a single image into a 360-degree panoramic video along a camera path you choose, then rebuilds it as a 3D scene you can move through. Built on Wan 2.1, with a smaller 5B variant based on Wan 2.2.

open source, MIT · needs 16GB VRAM minimum for the full pipeline; 720p video generation wants about 60GB, or about 19GB in low-VRAM mode; the 3D reconstruction step about 80GB

Marble (World Labs)

A commercial tool for creating and sharing 3D worlds in the browser.

hosted service · needs none, runs in the cloud

LightX2V

An inference engine that runs existing text-to-video, image-to-video and text-to-image models faster and in less memory than the reference code.

open source, Apache-2.0 · needs 14B models at 480p or 720p run on 8GB VRAM plus 16GB system RAM

LichtFeld Studio

A desktop application for training, editing, viewing and exporting Gaussian splat scenes, all in one program rather than a chain of scripts.

open source, GPL-3.0 (so anything you build on it must also be GPL) · needs NVIDIA GPU of compute capability 7.5 or newer (GTX 16-series, RTX 20-series and up) with driver 570+ and CUDA 12.8+; GTX 10-series, AMD and Intel GPUs are not supported

Krea 2 Technical Report

Krea's image generation models, released with downloadable weights in a standard and a fast version.

open weights

HY-World 2.0 (Tencent Hunyuan)

Tencent's system for generating a navigable 3D world from text, an image or a video, and exporting it as a mesh or Gaussian splat. The repo does not state the four-GPU requirement the video mentions, only that in multi-GPU mode you need at least as many input images as GPUs.

open weights under the Tencent HY-World 2.0 Community License: free below 1 million monthly active users, and the licence explicitly does not apply in the EU, the UK or South Korea

Brush

A Gaussian splat trainer and viewer that ships as a single binary with nothing to install, which is why the video reaches for it over the research code.

open source, Apache-2.0 · needs runs on Windows, macOS, Linux and Android with AMD, NVIDIA or Intel GPUs; the browser version needs Chrome 134+ on Windows or macOS

Apple SHARP Project Page

Apple's method for turning one photograph into a viewable 3D scene in under a second, with paper and code published.

open source

Apple SHARP Paper

Apple's paper on reconstructing a viewable 3D scene from a single photograph in under a second.

paper only

Apple SHARP

Apple's method for predicting a full 3D Gaussian scene from a single photo in under a second. The video's objection to its licence holds up: the code is permissively licensed but the released weights are for non-commercial research only.

open weights, non-commercial research only; the code is separately under Apple's own permissive licence · needs prediction runs on CPU, CUDA or Apple Silicon; rendering video out needs a CUDA GPU

2026-08-23 AI Search

SenseNova U1.5

An 8-billion-parameter reasoning model from SenseTime, distributed through ModelScope.

open weights

Qwen Video Edit

Edits video by written instruction, including long clips where each segment gets a different instruction.

research project

Ornith 1.5

A model family built to improve its own answers in a loop rather than in a single pass.

open source

Happy Shrimp

A music service that writes a complete song, lyrics and vocals included, from a one-line prompt.

hosted service · needs none, runs in the cloud

GeoWeaver

Builds a consistent 3D reconstruction from long video sequences without drifting off course.

research project

Gen 1.5

A robot model that picks up a new task from a few seconds of demonstration, or a few minutes of data.

hosted service

Evoke

Named in the video, but the project page it linked to no longer exists.

unavailable

Deepseek vision

DeepSeek's image understanding: send a picture with your question and it reads charts, screenshots and text.

paid service · needs none, runs in the cloud

Comfy MCP

Lets a coding assistant drive a ComfyUI install on your own machine, using your GPU, your models and your custom nodes.

open source · needs your own GPU

Bernini v2

ByteDance's video model that both generates and edits: a language model works out what the instruction means, then a renderer applies it, covering text- or image-to-video and prompt-based edits of existing footage.

open weights, Apache 2.0 · needs Hopper GPUs (H100/H800/H200) recommended

Audio8 TTS 0.1B

A very small text-to-speech model, 0.1 billion parameters, that clones a voice from a few seconds of audio; Chinese and English are supported, six other languages experimentally. The video noted only that the code and weights were released; the licence is revenue-capped rather than open.

open weights, Audio8 Community License v1.0 - free below US$2m annual revenue, paid licence above it · needs a CUDA-capable GPU is recommended; an INT8 CPU build needs about 0.4 GiB

4DAnyone

Rebuilds a moving 3D person from one ordinary handheld video, with no camera calibration.

research project

2026-08-16 AI Search

WorldClaw

Turns a single open-ended prompt into a 3D world you can walk through and edit.

research project

Sign to text

DeepMind's sign-language-to-text model, now driving features in shipping products.

hosted service

Scope

Gives an existing video model precise camera control by telling it which direction each part of the frame is being viewed from.

research project

Qwen 3.8 max

Alibaba's largest Qwen model, 2.4 trillion parameters with 95 billion active per token, aimed at coding, research and long multi-step work, with a 262k-token context. The video said Alibaba had made it open source; the weights are public but the licence is Alibaba's own "qwen3.8-max", not a standard open-source one.

open weights, under Alibaba's own qwen3.8-max licence

✓Qwen 3.8 27B

A mid-sized Qwen model that reads images and video as well as text, and can be told how much to think before answering.

open weights, Apache 2.0

Nemotron Lightning

A small open NVIDIA model plus a routing library, meant to run the same workload across a laptop, a workstation or a data centre.

open weights

Muse Glimmer

Meta's 30-billion-parameter model for agents that run continuously on your own machine, tuned for tool use and recovering from mistakes.

open weights, apache-2.0 · needs a single GPU

MiDasheng

Generates full audio scenes of variable length by treating sound the way a language model treats words.

research project

MatrAIx

Testing infrastructure that simulates a range of users against your product before you release it to real ones.

open source

Magi 2

Generates a video clip with its sound already in it, from text or from an image plus text; 114 billion parameters but only 6 billion active at a time, capped at 10-second clips up to 1080p.

open weights, Apache 2.0 · needs 8 NVIDIA Hopper GPUs, per the model card

LTX 2.5

Generates multi-shot video in one pass, edits real footage, and exports at cinema quality. Weights are downloadable.

open weights · needs runs on your own hardware

JoyAI video edit

Edits video from a plain-language instruction as the footage arrives, so it can work on a live camera feed rather than a finished file.

open source, Apache 2.0 · needs real time on one RTX 5090 (32 GB) at 840x480, or an RTX PRO 6000 for 720p

Index TTS 2.5

Clones a voice from a few seconds of audio and makes it read any text, with control over emotion, across Chinese, English, Japanese, Spanish and Arabic, including dubbing one language in a voice recorded in another. The video called it open source; the weights actually come under bilibili's own model licence, whose commercial terms are not stated on the page.

open weights, bilibili Model Use License Agreement · needs an NVIDIA GPU with roughly 6 GB of VRAM

Grok 4.6

xAI's assistant model, this version aimed at agents that run for a long time without supervision.

paid service · needs none, runs in the cloud

GPT Ultrafast

An OpenAI service tier that runs their model far faster on specialised chips, quoted at up to 750 words per second.

paid service · needs none, runs in the cloud

GLM 5.3

The announcement page for Zhipu's GLM 5.3 language model. The page returned no readable text to us; the downloadable weights are listed separately in this archive.

unclear

Gemini 3.7 Flash

Google's mid-tier Gemini model, positioned for everyday coding and agent work.

paid service · needs none, runs in the cloud

Dyna 2

A robot control model trained on over a million hours of ordinary human video, presented with evidence that it keeps improving with scale.

hosted service

Deepseek V4 0813

DeepSeek's flagship text and agent model, 1.7 trillion parameters, with three selectable levels of reasoning effort and a context window up to a million tokens.

open source, MIT

Deepseek harness

DeepSeek's own agent tool for driving its models from the command line or a local web page, the equivalent of Claude Code or Codex; built so that every capability is a plugin, and still a developer preview with breaking changes expected.

open source, MIT

Cactus Needle

A tiny 45-million-parameter model for calling tools and pulling structured data out of text, small enough to ship inside an app.

open weights · needs runs in about 28 MB of memory

2026-08-15 AI Search

Muscriptor

The link points at a publisher's Hugging Face account rather than a specific release, and the page could not be read without signing in.

unclear

Minimax Music for Comfy

Official instructions for running MiniMax Music 3 inside ComfyUI to produce complete songs up to five minutes long.

free to use

2026-08-12 AI Search

Smaller models

A scratchpad of unfinished MiniMax H3 parts from the person who converts most models for ComfyUI: a 4-bit weight format he says is still being tested, a smaller 8-bit video decoder that decodes about 1.5x faster, and a LoRA he says he is not sure has a use. Some pieces needed ComfyUI changes that had not been merged. The repo states no licence.

open weights

Realism lora

An add-on from fal that pushes MiniMax H3 towards believable footage of people, keeping skin texture and eye detail instead of smoothing them away. The trigger word is spelled r34l1sm, and the card suggests a strength of 0.6 to 1.0. The card says it inherits the base model's community licence.

open weights, not licensed in the EU, UK, South Korea or the US

MiniMax H3 Turbo Lora

A speed-up add-on that cuts MiniMax H3 from about 20 sampling passes to 4-8, roughly five times faster, while still producing video and matching stereo audio together. Its author calls it a preview with training still going, and recommends 6 to 8 passes rather than the bare minimum.

open weights, apache-2.0 · needs an 80 GB GPU is comfortable at the largest sizes; a low-VRAM switch and a CPU offload worth about 13 GB are offered for smaller cards

MiniMax H3 Turbo

A fixed four-step speed-up add-on for MiniMax H3 in ComfyUI, for text- and image-to-video, built against the full-size BF16 model rather than the shrunken ones. Its author is candid that four steps costs some fine detail and steadiness. The repo sets no licence of its own and points to the base model's terms.

open weights

Minimax h3 Turbo

LightX2V's own speed-up build of MiniMax H3, deployed as an 8-step first/last-frame adapter at 768p, offered both as downloadable weights and through the team's hosted studio.

open weights, apache-2.0

MiniMax H3 prompt guide (reference)

The companion guide for H3's reference mode, where you hand it existing pictures, clips or audio. It sets out a labelling scheme for each source and asks you to state how much of each should survive into the result, from fully preserved down to a weak hint, with video and audio numbered separately.

free to use

MiniMax H3 prompt guide (basic)

MiniMax's own guide to writing H3 prompts, which are not one-liners: it asks for three labelled blocks covering the picture and dialogue, the ambient sound, and the background music the characters cannot hear, with camera moves, named speakers and timestamps. Four worked examples cover the text, image and first/last-frame modes.

free to use

MiniMax H3 GGUF

Compressed GGUF conversions of the MiniMax H3 video model, 3-bit to 8-bit and roughly 16 to 36 GB a file, in both the first/last-frame and reference-driven variants, so it fits a smaller machine.

open weights, not licensed in the EU, UK, South Korea or the US

MiniMax H3 comfy

MiniMax H3 repacked into single files ComfyUI can load, plus a shrunken copy of LightX2V's speed-up LoRA that the card suggests running at 4 steps and 0.75 strength. The repo states no licence of its own.

open weights

MiniMax H3

MiniMax's own release of H3, a video generator that reads text, images, video and audio together and returns 4 to 15 second clips at up to 2K with native 32 kHz stereo sound, prompted reliably in 11 languages. Videos call it open source; the licence is not, and the card carries an application form only for the USA, EU, UK and South Korea.

open weights, not licensed in the EU, UK, South Korea or the US

Live preview

A tiny decoder that turns MiniMax H3's half-finished working data into a rough picture, so you can watch a video take shape in ComfyUI and cancel it early if it looks wrong. Its author describes it as quickly trained and not great, just better than the usual preview trick.

open weights, apache-2.0

Image workflow

Three loose files rather than a documented release, with an empty read-me: a ComfyUI workflow that coaxes still images out of the H3 video model by asking it for a single frame, an audio experiment, and a text file of example prompts for infographics.

free to use, under the MiniMax H3 licence

GGUFs

Unsloth's compressed conversions of the MiniMax H3 video model, from 2-bit up to 8-bit and roughly 6 to 20 GB a file, so it fits on a smaller machine.

open weights, not licensed in the EU, UK, South Korea or the US

Easy workflow

Collapses MiniMax H3's various modes — text to video, image to video, first/last frame, reference video and digital human — into a single ComfyUI node instead of a large hand-wired graph.

open source, mit

2026-08-05 AI Search

Workflows and files

Official ComfyUI instructions for MiniMax H3, covering text, image and reference-driven video with stereo sound.

free to use

Spectrum

A ComfyUI add-on that speeds up MiniMax H3 by predicting some of the model's intermediate steps instead of computing them — roughly half the expensive steps skipped, with the caveat that the result will not exactly match an unaccelerated run of the same seed.

open source, gpl-3.0

Sage attention wheels

Ready-built Windows installers for SageAttention, a faster attention routine that speeds up image and video models; the point of this fork is that you do not have to compile it yourself.

open source, apache-2.0 · needs Nvidia cards; the newest kernels need CUDA 12.8 or later with an RTX 40- or 50-series card, older cards fall back to slower kernels

AI models and tools worth knowing about

2026-11-01 AI Search

2026-10-28 Mickmumpitz

2026-10-25 AI Search

2026-10-23 AI Search

2026-10-20 AI Search

2026-10-18 AI Search

2026-10-11 AI Search

2026-10-07 Mickmumpitz

2026-10-06 AI Search

2026-10-04 AI Search

2026-09-27 AI Search

2026-09-24 AI Search

2026-09-22 AI Search

2026-09-20 AI Search

2026-09-18 AI Search

2026-09-15 AI Search

2026-09-13 AI Search

2026-09-11 Mickmumpitz

2026-09-10 AI Search

2026-09-07 AI Search

2026-09-06 AI Search

2026-09-03 AI Search

2026-08-30 AI Search

2026-08-25 Mickmumpitz

2026-08-23 AI Search

2026-08-16 AI Search

2026-08-15 AI Search

2026-08-12 AI Search

2026-08-05 AI Search

2026-08-02 AI Search

2026-07-26 AI Search