Open-weight models released since 2022: text, code, images, video, speech, 3D. For each one — what it does, where it is used, what hardware it needs and whether commercial use is allowed. I can deploy any of them on your server and fine-tune it for your task.
A dense (not MoE) model with 65B parameters, pretrained on 12 trillion tokens, with a context of up to 1 million tokens. It is built for programming and agents: by the authors' own test it scores 80.52 on Terminal-Bench 2.1 and beats much larger American open models (Reflection Beam 501B with 80.1, Meta Muse Spark Open 816B with 76.4, NVIDIA Nemotron 3 Ultra 550B with 56.4). The authors started from the Olmo 3 recipe and claim a 16 to 55 times gain in training efficiency on code. It was designed for NVIDIA graphics cards. BF16 and NVFP4 are published for the instruction-tuned version, plus a BF16 base. This is a preview release, and the figures are the authors' own, not independent.
an agent for programming in the terminal and a repository
long tasks with a context of hundreds of thousands of tokens
Very large Moonshot MoE models for agentic work. K3 (2.8 trillion parameters) was the largest open model at release, with up to 1M tokens of context and image understanding; K2.7-Code is built for programming. On November 4 the weights of K3.1 Pro were opened (2.8 trillion parameters, 104B active, 1M tokens of context, images): a refinement of K3 for programming, 3D and financial research; by the tests of Moonshot itself, Terminal-Bench 4.0 is 65.7%. The weights are in MXFP4, about 1.56 TB, run through vLLM and SGLang, and need a serious cluster.
Multi-step agents: search, data collection, reports
Sber's Russian family of image and video generation models. Understands Russian-language prompts and Russian cultural context well; released under MIT. The new Kandinsky S2V 1.0 (Avatar, MIT) animates a person from a single frame to a speech recording: the whole figure moves, with expressions and gestures, in clips of up to a minute. Pro (HD) and Lite (SD, runs on a home graphics card) versions; it was trained on more than a million fragments in 95 languages.
Indian models focused on 22 languages of India. Sarvam 30B and 105B (2026) are MoE models with strong reasoning and agent skills. In November Sarvam-Prism came out (Apache 2.0, based on Sarvam-105B): a decision model that, given a context and questions with fixed answer options, returns the probability of each in one pass and writes no text. It needs a graphics card with FP8 and about 120 GB of memory (B200 or H200); a paid API exists too.
A compact 4B model that operates a computer by itself: it looks at a screenshot and clicks, and when that is more efficient it switches to the command line and edits files with code. Both roles are in one model, with no external orchestrator. On the OSWorld-Verified test (50 steps) it scores 66.0%, which the authors say is the best result among 4B models, and 87.5% on operating-system tasks. It was trained from the intermediate Mage-VL in five stages, with a set of almost 294 thousand synthetic tasks over 6,915 skills. The context for training and running is 32,768 tokens, and the architecture holds up to 262,144. It suits office automation and assistants on your own computer. The figures come from the authors.
an agent that works in applications through screenshots and the command line
automating office tasks on your own computer without a cloud
NVIDIA models for agents and reasoning, optimized to run fast on its GPUs. Nemotron 3 is a Mamba and MoE hybrid from 4B to 550B; Nano Omni handles video, audio and images (English only). The new Nemotron 3.5 Super VL (120B, 12B active, up to 1M-token context) reads text, images and video up to two minutes, with a focus on documents and agents; the compressed Bolt (74B, 9B active) fits into 24 GB of video memory. Super VL languages: English, Spanish, French, German, Italian, Japanese. Dory 1.0 (4B, up to 256K-token context, English) runs the same block of layers several times, and the number of passes can be changed per task: more passes give more accuracy without a new model and without long reasoning in the text.
lodestones (independent developer) · not disclosed
A community model retrained from FLUX.1-schnell with a simplified architecture. No style censorship; popular as a base for fine-tuning. In October Chroma-H3 (lodestones) came out: an image fine-tune of the 33B MiniMax H3 video model with speed-up adapters for 4 and 8 steps; its license comes from MiniMax H3.
From a single picture of an object it makes a three-dimensional model in the form of Gaussian splats (the 3DGS .ply format). It takes a render, a generated image or a photo, removes the background itself, frames the object and fills in sides that the shot does not show, up to a million Gaussians. It runs on an ordinary graphics card with 7.2 GB of memory, in two seconds according to the authors. There are modes for renders and for photographs. The result opens in any splat viewer. The weights are Apache 2.0, but running it needs two gated encoders (DINOv3 and the VAE from FLUX.2-dev) with their own terms, which affects commercial use.
quick 3D previews of products from one photo for cards and catalogs
A 33B model that the authors call the senior teacher for future lightweight agents: they plan to train small models (for example 3B) on it. It is strong in programming and office tasks: on the JobBench test across 35 professions it beats the baseline in most, according to the authors. Context up to 512 thousand tokens, reasoning can be switched on and off, and tool calling works best in the XML format. The authors compare it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma 4 31B and say it leads on most tests; the figures are their own. A base version is published too.
an agent for programming and working with tools
office tasks by profession: reports, tables, document analysis
A Pika video model built on the open MiniMax H3 weights: it makes video with stereo sound at once in six steps and takes up to 12 reference materials (images, video, audio) that set the shot direction. Two adapters: FL2VA for text and the first and last frames, Ref2VA for reference materials. Locally it runs at 480p and 768p, while 1080p, 2K and 4K, audio stems and prompt enhancement are available only through the paid Pika API. It suits multi-shot stories, motion graphics and product films. The license inherits the MiniMax H3 terms and does not apply in the EU, the UK, South Korea and the USA, which has to be checked before use. Running it needs server-grade graphics cards.
ads and product films with ready sound
animation from reference images and video, multi-shot stories
Open weights of MiniMax's Hailuo video model. A large 33B model that makes video from text and images, but needs several server GPUs. The license does not apply in the European Union, the United Kingdom, South Korea and the USA. The October fine-tune of these weights is Pika H3 Flex.
Fast NVIDIA speech recognition models, including streaming ones for real-time use. Parakeet TDT v3 and Nemotron 3.5 ASR understand Russian. The new Nemotron 3.5 Transcribe (34B, 4.3B active, OpenMDW 1.1) transcribes recordings of up to an hour in 40 languages, translates speech, splits it by speaker, puts word timestamps and answers questions about the content (questions in English only); all tasks are available for Russian. Ready files only, live streaming is not supported. The figures are the authors: 5.33% word errors on English in the open Hugging Face leaderboard.
A "living world" video model: it builds the world segment by segment, and you steer it with text and a camera trajectory. At each step you can add an object, change the style or scene, describe an action or turn the camera, and the memory of the world carries over between segments. Output is 720 by 1280 at 24 frames per second, and a segment takes 8 steps. The standard configuration is meant for several graphics cards, but there is a variant for a single RTX 4090 with INT8 weights and part of the model offloaded to the CPU. For narrow areas such as robotics a light LoRA tuning can be attached. Together with the model came the Hy-Nomas-Bench set of 280 test cases; the technical report and tests of the model itself are promised later.
interactive video worlds and game-scene prototypes from a description
test videos for robotics with LoRA tuning
experiments with a controllable camera and scene consistency
Sizes
not stated (54 MMDiT blocks, a Qwen2.5-VL-7B text encoder)
An Ant Group family: Ling for standard models, Ring for reasoning ones. There are trillion-parameter flagships and the efficient Ling-3.0-tiny, which needs only 1.3 billion active parameters. The new Ling-3.1-flash (560B, 25B active, 1M-token context) targets agents, office tasks and development. The improved Ling-3.1-flash-1019 (MIT) raised the Artificial Analysis index from 41 to 44 by the authors estimate.
Small decision models: from text, an image or a voice clip they answer yes or no, pick an option or give a rating in one pass, without writing text. d1-3B sees images, d1-omni-600M also hears speech up to 30 seconds. Russian is listed. The open d1-2B (text, pictures and video, based on LFM3-VL-2B) and d1-Omni-2B (sound as well) decide in milliseconds, 5.7 ms on an H100 by the authors measurements; besides choosing an answer they pull out text, numbers and dates with their positions and find objects in a picture. It is not a chat model.
routing requests and tickets to teams
checking a photo for defects or violations against a list of questions
Models with a new architecture for on-device use: fast on a regular CPU and on phones. Versions for data extraction, RAG and tools, plus LFM2.5-VL for images and voice LFM2.5-Audio. The largest, LFM2.5-24B-A2B (2.45B active, 131K-token context, Russian in the language list), always reasons before answering and targets fast on-device agents. In October the LFM3 generation came out on a new tri-hybrid architecture: LFM3-1.7B for local agents (256 thousand tokens of context, about 1 GB of memory at 4 bits, pre-trained on 22 trillion tokens), LFM3-VL-2B for images, LFM3-Audio-2B for speech and LFM3-Omni-2B, which understands text, images, video and sound and answers in text and speech. The authors do not recommend it for complex code and multi-step reasoning.
The first MiniCPM model built only for documents: it reads text, tables, formulas, charts and seals, and pulls the needed fields out of papers. Just 1.3B. By the authors' tests, 93.68 on Real5-OmniDocBench (document parsing in real conditions: scans, photos, skewed and recaptured pages) and 54.7 on OCRBench v2. Chinese and English are stated, Russian is not. An auxiliary DSpark model is published for it that speeds up generation: according to the authors processing is 55% faster, and the speed is about 1.5 times that of PaddleOCR-VL-1.6 and 3.6 times that of MinerU2.5 Pro. It runs in vLLM, SGLang, llama.cpp and Ollama, with compressed GGUF, AWQ and GPTQ versions.
parsing scans and photos of documents into markup: text, tables, formulas
extracting fields from invoices, contracts and forms in Chinese and English
fast processing of a document flow on a small server
A compact 4B model to run right on a device, focused on agents: tool calling, multi-step planning, programming, long context. Version 2.6 came about a month after 2.5 with the same architecture and improved only through further training. The reasoning budget is adjustable: none, low, high. The authors state that it holds its own against models of around 9B. Languages on the card: English and Chinese, Russian is not stated. Running it needs the model's own code in transformers. The figures come from the authors. At the end of October a 1.7B size was added (INT4, INT8, FP8).
an agent with tool calling on a home computer or laptop
a local programmer assistant that does not need a big server
AI built into an app with a choice of how much the model thinks
An Ant Group family for finding elements on screen and completing tasks in phone and computer interfaces. UI-Venus-2 was specifically trained to refuse dangerous actions. In October LBS-27B came out (Apache 2.0, based on Qwen3.8-27B): it learns to pick and zoom into the needed parts of an image on interface markup, and carries this skill over to ordinary questions about pictures.
Baidu's OCR model building on DeepSeek-OCR ideas: processes multi-page documents and PDFs in a single pass and outputs structured text. Claimed to be multilingual, but the language list is not published. On October 27 Unlimited-OCR-2 (MIT, about 3.3B) came out: according to the authors it parses documents of unlimited length (the SR-SWA technique), and the MarathonDocBench set for very long documents was published with it.
Converting multi-page PDFs and scans to text and Markdown
vLLM Semantic Router project · international project
An open family of models for an AI gateway: they answer a request with probabilities without generating text. They detect topic and urgency, jailbreak attempts, harmful requests, personal data and whether a tool is needed. Small versions run on a CPU, Vela understands 17 languages including Russian. The new d3 (27B) also reads several images and videos, and d3-turbo is, by the authors' own estimate, 4 points more accurate on text. The newest in the line, d3-ultra, came out on October 24.
routing requests between models and teams
checking a request for a jailbreak attempt or harmful content before the main model
An open 14.8B language model for working with DNA, built on Qwen3-14B. You send a raw sequence right in an ordinary chat, and it answers in English: it judges whether a mutation is benign or pathogenic, finds regions that control gene activity, assigns a sequence to an organism group, predicts the next bases and explains its reasoning. DNA and text go in one stream, with DNA encoded in groups of six bases. By the authors' tests, on five sets it scores 66.7% without reasoning, 27 points above Qwen3-14B, and costs about 1,800 times less than GPT-6.1 Sol at a close result. It fits on a single 80 GB card, so sensitive data can stay with you. Viral data was excluded from training because of the risks. The conclusions are computed and do not replace laboratory verification.
assessing DNA variants: benign or pathogenic
finding regions that control gene activity
a research assistant for biologists on your own server
A family of open speech models from Ant Group: a dense 0.5B based on Qwen2.5-0.5B and the MoE version tiny (7.9B parameters, about 1.3B active, based on Ling V3). They clone a voice from a short sample, create a voice from a description in words, and control delivery, emotion, speed, pitch and volume. There are more than 100 built-in Chinese voices, five English ones and 14 inline tags for a breath, cough, laugh, sigh and hesitation. Languages: Chinese, English, German, Spanish, French, Korean, Japanese and the Sichuanese and Cantonese dialects, no Russian. Sample rate is 44.1 kHz. Running it needs the code from the GitHub repository, and the Hugging Face page only distributes the weights. There is a Comfy-Org build for ComfyUI.
voicing videos and podcasts in Chinese and English with emotion control
cloning a voice from a short sample
creating a voice from a description for games and apps
Large MoE models with very long context (up to 1M tokens for Text-01 and M3). M3 is multimodal and understands images. In October the open weights of M3.1-Flash came out (about 428B, 23B active, up to 1M tokens of context, text, images and video): a model for programming, agents and computer use, sped up by sparse attention. MiniMax Community License: free for non-commercial use; commercial use needs the caption "Built with MiniMax M3.1-Flash" and a one-time notice to MiniMax, and a separate permission when the product earns $20 million a year or more; military use is prohibited. The benchmark figures come from MiniMax itself. Licenses differ greatly from version to version.
Analysis of large document archives in a single request
A diffusion model that creates three-dimensional body animation from a text description and set conditions: poses, positions of hands and feet, a route and points to pass through. There are versions for the SOMA skeleton (30 joints), the Unitree G1 robot (34 joints) and the SMPL-X human (22 joints). Version 1.2 is trained on a corrected set: in 1.1 some test recordings leaked into training, so the new version is the one to compare against other models. According to the authors, foot skating is reduced and text following is slightly weaker. The data of some variants is proprietary (Bones Rigplay), while the open variants are trained on Bones-SEED.
animating characters for games and video from a description
motions for humanoid robots (the Unitree G1 skeleton)
"World" models for robots and self-driving vehicles: they generate realistic video of physical scenes and predict actions. Cosmos 3 combines understanding, generation and control. In October the surgical versions came out: Cosmos3-H-Embodied-Nano (16B, one model for predicting video, actions and policy on surgical robots) and Cosmos3-H-Dreams (a real-time world for simulating operations with the Versius robot). They are under the NVIDIA Open Model License.
Synthetic video for training robots and self-driving vehicles
NVIDIA embeddings for search and RAG. Nemotron-3-Embed, released in 2026, is under the permissive OpenMDW license and works in many languages. In October Nemotron-3.5-Embed (1B and 8B, OpenMDW 1.1) came out: it finds documents by text and by the image of a page with tables, charts and infographics; for text it was tested in 34 languages, Russian among them. The 8B is built on Ministral 3.
A model from the Korean KRAFTON (the publisher of PUBG) with 30B parameters, 4B active, trained from scratch. It is strong in Korean conversation and agentic tasks: it can reason and call tools, and works with Hermes Agent, Pi and OpenCode. Training languages: Korean, English, Japanese, Chinese, Spanish, French, Thai, and Russian is not among them. There are compressed MXFP8 and NVFP4 versions and an auxiliary model for faster generation. This is a trial release (preview), and the specifications may change before the final one. The weights are issued on request on Hugging Face, and the benchmark figures are the authors' own.
An open expert model (39.8B parameters, 2.5B active per token) for coding agents: it digs into someone else's repository, fixes bugs, writes functions and tests, and can reason and call tools over several steps. Trained from scratch, 262 thousand token context, text only. It weighs about 80 GB, and the authors run it on a single B300 or H200 card through SGLang. By the authors' own tests: Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2%. Part of the comparisons come from other developers' own reports, and the test conditions differ between sources.
an agent for fixing bugs and extending a repository in the terminal
writing tests and new functions from a description
Contrastive-LM (per the announcement: Stanford together with NVIDIA) · USA
An open fast-choice model for agents: it takes a description of a situation and a set of options and returns the probability of each without generating text. It can answer yes or no, pick from a list and give a score on a scale. Actions and state are encoded separately, so action vectors can be computed in advance: by the authors' measurements, with a choice among about a thousand options it is up to 13 times faster than the Jev analogue. After fine-tuning it works as a verifier that picks the best of several solutions of a coding agent (DeepSWE 81.6%, Terminal-Bench 2.1 87.6% by their measurements). It needs Linux, an NVIDIA card and Qwen3-8B in vLLM, and the card language is English.
fast choice of an agent's next action from given options
checking and ranking several solutions of a coding agent
routing requests: department, urgency, a score on a scale
Sizes
8B (a frozen Qwen3-8B plus two small projection heads)
A family of Bria image models: Fibo 2.0 dev draws from text, Fibo 2.0 edit turbo edits finished images from a description. According to the authors' paper, the model was trained on very detailed structured descriptions so that it follows a long brief better. The weights are free only for non-commercial work: advertising and sales need a separate Bria license. There are ready files for ComfyUI, including ones compressed to int8. The languages of prompts are not stated on the card.
creating pictures from a detailed description for tests and prototypes
editing finished pictures with a text command
research and personal projects without commerce
Sizes
not stated (the bundle includes a Qwen3-VL 4B text encoder)
Models for agentic development: they build their own plan and scaffolding for a task and execute it in the terminal. Fine-tuned from Qwen 3.5 and Gemma 4; work with Claude Code, OpenHands and similar tools. In October Ornith-1.5-122B (MoE, MIT) came out: according to the authors it is stronger and faster than the previous ones, writes shorter and is meant for a single DGX-class card; compressed FP8, NVFP4, GGUF and MLX versions exist. The figures come from the authors.
A fast climate model emulator: simulates the atmosphere years and decades ahead on a single GPU. Coupled with an ocean model (SamudrACE) for long-term scenarios. The new HiRO-ACEv2 adds a diffusion model that sharpens the coarse 100 km grid to 3 km: precipitation, wind and pressure every 6 hours for decades ahead, worldwide or for a chosen region. By the estimate of Ai2, global 3 km data come out about 300 times cheaper than with the X-SHiELD model. For now it reproduces only the X-SHiELD climate, not real weather from the archive.
Decades-long climate scenarios to assess long-term asset risks
Large-scale what-if runs on temperature and precipitation
Preparing data for crop yield and energy demand models
A 12B model that rewrites drafts written by AI so that they read like a person wrote them: emails, essays, reports, forum posts. It is trained to keep every number, unit, date, name and quote and to add nothing. It runs on your own computer; there is a ready app for Mac and Windows that works offline, and GGUF files of different compression for llama.cpp. Stated languages: English and Chinese, Russian is not stated. By the author's measurements, the Originality.ai detector judged 95% of the English rewrites human. The result still has to be read through, especially numbers and names.
rewriting drafts of emails and reports so that they sound more natural
editing texts after AI without losing numbers, dates and quotes
local processing of texts that must not be sent to the cloud
A model that runs the long-term memory of an AI agent: it decides what from a chat, work tasks or code is worth remembering, in what form, and how new information changes what is already stored. It replaces the general-purpose model inside the memory systems Mem0, MemOS and Tencent's own Hy-Memory. Languages on the card: English and Chinese. The benchmark figures are Tencent's own.
an assistant's memory about a customer between conversations
remembering the course of work on a project and code
updating stored facts when new data arrives instead of duplicates
A text-to-image model. It takes an ordinary description, sets of tags and a JSON layout in which you can set where a caption stands and how objects are arranged. It renders text in English and Chinese well. Base and fine-tuned (SFT) versions exist, and it plugs into ComfyUI. It uses a text encoder based on Qwen3.5-9B and the FLUX.2 decoder. The community is already fine-tuning it: Kromoss (lodestones) removes the blurriness of the base version, is a drop-in replacement for SFT and comes with an 8-step speed-up LoRA.
Qwen3.8-27B squeezed to 2 bits into a single 7.9 GB GGUF file instead of 54 GB, with the focus on keeping tool calling intact. By the authors' measurements (their Underdog Bench), tool calling is 88 against 84 for the full model, parallel calls 42 against 35, and on average 96% of the full model's quality across nine tests. It runs on plain llama.cpp and Ollama, thinking mode can be switched on and off, and it understands images with a separate vision file (about 0.6 to 0.9 GB). Language stated on the card: English.
an agent with tool calling on a home graphics card
A tiny OCR model that runs on an ordinary CPU with no graphics card and no PyTorch: it uses ONNX Runtime, takes about 230 ms per page on four cores and reads up to 49 pages per second. It turns PDFs and images into markdown, plain text or JSON with boxes for lines and blocks (headings, lists, tables, captions) and reading order. In digital PDFs it takes the existing text layer and recognises only what is missing. Russian is among the stated languages. For handwriting, formulas and complex tables the authors point to their larger Chandra. The speed and accuracy figures are the authors' own.
fast bulk conversion of PDFs and scans to markdown on a server with no graphics card
search over document archives: text with boxes and reading order
pre-marking pages before a heavy model
Sizes
about 1M parameters (the models weigh about 18 MB)
An open model from the American company Reflection with 501B parameters (23B active, a 1 million token context), text only, mainly English. Stated as economical in reasoning: according to the authors it needs about three times fewer tokens for thinking at comparable quality. Strong in programming, tool calling and science. Weights in BF16, FP8 and NVFP4. The benchmark figures are the authors' own.
an agent for solving tasks in a repository and a terminal
long tasks with tools and search (via MCP)
solving scientific and mathematical problems with reasoning
DeepSeek's flagship line: from the first 7B/67B to V4-Pro with 1.6 trillion parameters. Closed-model quality under an open MIT license; V4-Flash-Vision-Exp and V4.1-Flash understand images, context up to 1M tokens. The fresh V4.1-Flash-1017 answers shorter and faster than the previous Flash: by the measurements of DeepSeek itself, 77.0 vs 74.2 on DeepSWE v1.1 and 42.3 vs 31.2 on Terminal-Bench 4.0.
Meta's largest open model: 816B parameters, 42B active per token, a 1 million token context. It understands text, images, video and audio and answers in text. Built for programming and long agentic tasks. Trained on data in more than 100 languages. Access to the weights on Hugging Face is granted manually.
an agent for working with code and long tasks with tool calling
analyzing screenshots, charts, documents, video and recordings
A family of models for forecasting time series with no task-specific training: from the history of sales, demand or load it gives a forecast with intervals (21 quantiles). It takes up to 32 thousand points of history and forecasts up to 1024 steps at a time. By Mistral's measurements, the best zero-shot model on GIFT-Eval, TIME and fev-bench. Runs on a CPU.
forecasting sales and demand from history
forecasting load on a warehouse, a till, a call center
assessing the risk of stock shortage or surplus from forecast intervals
A small 1B reranker from NVIDIA. The vl version also takes document pages as images, not just text. The new 8B (Nemotron-3.5-Rerank) was evaluated on 34 languages, including Russian, and accepts text, page images or both, including charts and tables.
Reordering passages before an AI assistant answers
Image generation from the creators of Stable Diffusion. Renders text in images well and keeps the composition. The new FLUX.3 Image Klein (5.6B) targets ordinary graphics cards: up to 4K at once, regional editing, up to ten references; access to the weights is granted automatically after signing in on Hugging Face.
Embeddings for mixed-modality search: text, images and pages with text and a picture together land in one space, and a single text query can rank them in one shared list. Built on research showing that ordinary models prefer text over images. Language on the card: English.
searching a catalog where some items are described by text and some by a picture
Quickly identifies the language and dialect of text in Arabic script: Modern Standard Arabic, all dialects and 51 classes overall. It tells apart close languages in the same script (Persian, Urdu, Pashto and others) so that a text is not called Arabic just because of its letters. Runs on a CPU.
detecting the dialect in customer requests in Arabic
sorting texts by language in Arabic script
preparing data for training and search
Sizes
a fastText model with calibration, parameter count not stated
Fudan University, Tencent Hunyuan and Zhejiang University · China
Creates video together with sound: from a first frame and a description you get a clip with speech, effects and music. Native generation at 720p, 1080p and 2K is stated. In the third-party FreeVideo build (1280 by 720, 8.5 seconds) it runs on one RTX 30, 40 or 50 series card with 12 GB or more, but it will not be fast.
Generates images directly in pixels, without a compressed latent space. After fine-tuning, the same weights estimate depth from a single photo and restore and upscale images 4 times. Prompts are in English, the default size is 1024 by 1024.
One compact model turns a document page, chart, flowchart, geometry figure, audio or video into structured JSON: with block coordinates, tables, formulas, speech transcription and timestamps.
parsing scans and PDFs with tables and formulas
turning charts and flowcharts into a table or diagram
Shanghai Artificial Intelligence Laboratory · China
An agent model for long tasks: searching and gathering evidence, research, development, preparing reports and presentations, analyzing vulnerabilities in an authorized environment. A preview based on GLM-5.2, 256K-token context, text only. An FP8 version exists.
deep search and research on a topic with a verifiable result
developing programs and preparing reports and presentations
checking and fixing vulnerabilities in an authorized environment
Finds the right pages in documents as pictures, without turning them into text: matches a query against how the page looks. Version 8B scores 66.75 on ViDoRe V3. The index can be compressed to 32 vectors per page.
searching scans, slides and reports without text recognition
A compact agent for working with repository code, based on Qwen3.5-4B: searches the project, edits files, runs tests and prepares a patch. Trained only with reinforcement learning on about 1500 synthetic tasks. Works through the light Leaf harness with five tools.
fixing bugs in a repository from a task description
Shanghai Artificial Intelligence Laboratory · China
A decision model based on Qwen3.5-4B: from text and images it answers all questions about a situation with probabilities in one pass, without generating text. Suits routing requests and rating urgency.
A pretrained model for tables: classification and regression from in-context examples, with no task-specific training. You supply labeled rows and the rows that need a forecast and get class probabilities.
A language model for computed tomography: looks at a whole chest or abdominal CT volume, not separate slices, writes structured reports and answers questions with reasoning. English only.
Parses a whole document page: turns it into Markdown with LaTeX formulas and HTML tables. The 4B version scores 95.38 on OmniDocBench v1.6. Languages on the card: English and Chinese.
turning scans and PDFs into Markdown for search and RAG
An open agent for operating a desktop: looks at live screenshots, reasons and outputs clicks and keyboard input. You can show a workflow once and the agent carries it over to a new task. An external environment is needed to execute the actions.
Universal embeddings based on Qwen3.5: they turn text, images, video and visual documents into one vector of 4096 numbers. Audio is not supported. Languages on the card: Chinese and English.
searching a product catalog by picture or description
MIT CSAIL, Massachusetts General Hospital and Harvard Medical School · USA
Finds 60 anatomical structures on an X-ray: ribs, vertebrae, pelvis and limb bones. They can be used to measure, for example, the curvature angle of the spine automatically. It does not look for pathologies and does not tell right from left.
labeling anatomy on X-ray images
measuring the spine curvature angle
matching fluoroscopy frames with CT for navigation
One model for developers: reads documents and tables, transcribes speech with speaker labels, finds objects in an image and answers strictly by a JSON schema you supply. Inside is a reasoning core and a set of narrow specialist models.
parsing documents and tables with text coordinates
transcribing recordings with speaker labels
extracting data from documents into a given JSON schema
Sizes
a model that needs 80 GB of video memory (the core is based on Qwen3.8-27B)
Hardware
from: 1 GPU
Commercial use with conditionsDetails
No models match these filters. Reset them or describe your task — I will pick one for you.
