Text2026
Ricursive Intelligence · USA
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
- your own tuning on top of an open base model
- Sizes
- 65B (dense)
- Hardware
- from: Cluster
Image + text2026
Microsoft · USA
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
- a prototype assistant that operates the desktop
- Sizes
- 4B
- Hardware
- from: Laptop
Text2026
Nanbeige · China
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 base model for your own fine-tuning
- Sizes
- 33B
- Hardware
- from: 1 GPU
Text2026
XHToken (the Spark project) · not stated
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
- Sizes
- 1.7B и 4B
- Hardware
- from: Laptop
Computer-use agentsGGUF2025–2026
Ant Group (inclusionAI) · China
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.
- Automating actions in mobile apps
- Filling in forms in web interfaces
- UI autotests
- Sizes
- 2B – 72B
- Hardware
- from: Laptop
Text2026
KRAFTON AI · South Korea
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.
- a Korean-language assistant and customer support
- an agent with tool calling on your own server
- game and media projects for a Korean audience
- Sizes
- 30B-A4B
- Hardware
- from: 1 GPU
Code2026
Recursive · not stated
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
- a local developer assistant on your own server
- Sizes
- 39.8B-A2.5B
- Hardware
- from: 1 GPU
Text analysis2026
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)
- Hardware
- from: 1 GPU
CodeOllama2026
DeepReinforce · not disclosed
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 developer agent in the terminal
- Fixing bugs from a task description
- Understanding and extending a large repository
- Sizes
- 9B – 397B
- Hardware
- from: 1 GPU
Text analysis2026
Tencent Hunyuan · China
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
- Sizes
- 30B-A3B
- Hardware
- from: 1 GPU
TextOllama2026
ConwayResearch (Underdog) · not stated
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 local coding assistant with functions
- analysing pictures with the vision file attached
- Sizes
- 27B (7.9 GB file)
- Hardware
- from: Laptop
Text2026
Reflection AI · USA
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
- Sizes
- 501B-A23B
- Hardware
- from: Cluster
Text2026
Meta Superintelligence Labs · USA
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
- generating training data for other models
- Sizes
- 816B-A42B
- Hardware
- from: Cluster
Text2026
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
- Sizes
- 744B (MoE, based on GLM-5.2)
- Hardware
- from: Cluster
Code2026
Microsoft · USA
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
- navigating an unfamiliar codebase
- preparing patches with test runs
- Sizes
- 4B
- Hardware
- from: Laptop
Computer-use agents2026
Tencent · China
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.
- automating work in programs with no API
- repeating a workflow from one shown example
- testing interfaces
- Sizes
- 27B
- Hardware
- from: 1 GPU
Computer-use agents2026
Beijing Academy of Artificial Intelligence (BAAI) · China
An agent model of 27B (based on Qwen3.8-27B) for long tasks: it proposes a solution, measures the result, analyzes errors and revises the solution over several rounds. A 262K-token context. Stated as a research model.
- iterative improvement of code and algorithms
- machine learning tasks with a verifiable result
- deep search and gathering information on a topic
- Sizes
- 27B
- Hardware
- from: 1 GPU
Text analysis2026
AutoTrust AI · Singapore
Open decision models: for yes/no, pick-one and rate-0-to-5 questions they answer with probabilities in one pass, and can also reason when needed. Not affiliated with the closed Jev from TypeSafe.
- routing requests to the right team
- checking an agent action before it runs
- rating urgency and tone of messages
- Sizes
- 9B – 35B
- Hardware
- from: 1 GPU
Text analysisOllama2026
Cloudflare · USA
Decision models: they do not write text but read a situation (text, JSON, an image, video frames, and for Clef-Omni also audio) and return in one pass the probability of each answer to typed questions: yes or no, pick one, rate on a scale. The agent gets ready numbers instead of parsing free text.
- sorting requests and tickets into categories with a probability
- checking a document or photo against a list of questions (is everything stated, is there a violation)
- choosing the next step of an agent without parsing a text answer
- Sizes
- 27B, 9B (Clef-flash) and 30B-A3B (Clef-Omni)
- Hardware
- from: Laptop
TextGGUF2025–2026
Renmin University of China (GSAI) and Ant Group (inclusionAI) · China
Diffusion language models: text is written in blocks and then refined rather than word by word, which speeds up generation. LLaDA2.2 can edit what it has written and targets agents. LLaDA-Image is a separate product.
- Fast generation of code and text
- Agent scenarios with long context
- Research into alternatives to standard LLMs
- Sizes
- 8B – 100B (MoE)
- Hardware
- from: 1 GPU
Computer-use agents2025–2026
XLANG Lab (University of Hong Kong) · China
Fully open desktop agents: weights, data and training code. They work on Windows, macOS and Linux; the latest Qwen-CUA controls a computer with ordinary clicks and keystrokes.
- Working in desktop software without an API
- Moving data between systems
- Running user scenarios for tests
- Sizes
- 7B – about 400B (MoE)
- Hardware
- from: 1 GPU
Computer-use agentsGGUF2025–2026
Microsoft · USA
Small Microsoft models for working in the browser: they look at the page and click, type and scroll. Designed to run directly on a work computer without the cloud.
- Filling in web forms and applications
- Collecting data from web portals without an API
- Checking websites against scenarios
- Sizes
- 4B – 27B
- Hardware
- from: Laptop
Rerankers2026
Tencent · China
A pair of small Tencent models based on Qwen3 that pick the right skill for an AI agent for a given request: the embedding model finds candidates, the reranker chooses the best one.
- Choosing a tool or skill for an AI agent
- Routing requests between bot scenarios
- Search across a catalog of internal tools
- Sizes
- 0.6B
- Hardware
- from: Laptop
Computer-use agentsGGUF2025–2026
H Company · France
A French model family for controlling a browser and computer: precisely finds the right element on screen and handles multi-step tasks. The latest Holo3 and 3.1 are open under Apache 2.0.
- Working in web portals and legacy software without an API
- Filling in forms and applications
- Testing interfaces against scenarios
- Sizes
- 0.8B – 235B-A22B
- Hardware
- from: Laptop
Computer-use agents2024–2026
Show Lab (National University of Singapore) · Singapore
A lightweight model for working with interfaces: finds buttons and fields by description and performs actions on the web and on a phone. ShowUI-π can drag with the mouse.
- Clicking and filling in forms from a task description
- Web UI autotests
- An assistant on a low-end computer without the cloud
- Sizes
- 2B (ShowUI), about 500M (ShowUI-π)
- Hardware
- from: Laptop
Computer-use agents2025–2026
Alibaba (Tongyi Lab, X-PLUG) · China
Models for controlling phones and computers from the Mobile-Agent project: they work with Android, Windows, macOS and the browser; version 1.5 has a reasoning mode.
- Automating actions in mobile apps
- Working in desktop software without an API
- Testing apps against scenarios
- Sizes
- 2B – 32B
- Hardware
- from: Laptop
Computer-use agentsGGUF2026
Meituan · China
Meituan's computer-control agent, trained on a large number of simulated tasks in desktop software. It outputs clicks and keyboard input.
- Working in office and legacy software without an API
- Moving data between systems
- Running test scenarios
- Sizes
- 8B – 32B
- Hardware
- from: 1 GPU
Computer-use agents2023–2025
Zhipu AI (Z.ai) and Tsinghua University · China
One of the first open models for controlling an interface from a screenshot; its successor, AutoGLM-Phone, works in Android smartphone apps.
- Automating actions in mobile apps
- Working in web interfaces without an API
- Testing apps against scenarios
- Sizes
- 9B – 18B
- Hardware
- from: 1 GPU
Computer-use agentsGGUF2025
Alibaba (Tongyi-MAI) · China
Compact Alibaba models for working in smartphone and computer interfaces: they find elements and complete multi-step tasks. The agent acts on the user's behalf inside their apps, so access rights and limits are set before launch. The follow-up line, Qwen-UI-Agent, was announced in July 2026, but its weights have not been released.
- Automating actions in mobile apps
- Working in software without an API
- UI autotests
- Sizes
- 2B – 8B
- Hardware
- from: Laptop
Computer-use agentsGGUF2025
ByteDance Seed · China
A model that looks at a screenshot and controls the mouse and keyboard itself: clicks, fills in fields, navigates menus. The first generation and 1.5-7B are open; UI-TARS-2 weights were not released.
- Working in legacy software without an API
- Filling in forms and moving data between systems
- UI autotests from plain-language scenarios
- Sizes
- 2B – 72B
- Hardware
- from: Laptop
Computer-use agents2024–2025
Salesforce · USA
Salesforce models for function calling and agents: they pick the right tool and fill in its parameters. Strong on benchmarks, but the license is non-commercial.
- Calling APIs and internal services on user request
- Multi-step agents with several tools
- Comparing approaches before choosing a commercial model
- Sizes
- 1B – 8x22B
- Hardware
- from: Laptop
Computer-use agentsGGUF2025
Microsoft Research · USA
An agent model that plans actions both in an interface (buttons on screen) and for a robot (arm movements). For now more of a research base than a finished product.
- Pilots in interface control
- Research projects spanning screens and robotics
- Analyzing screenshots with an action plan
- Sizes
- 8B
- Hardware
- from: 1 GPU
Computer-use agents2024–2025
Microsoft · USA
Breaks a screenshot down into buttons, fields and icons with labels so a regular language model can understand and control the screen. It does not click itself; it serves as the agent's eyes.
- Mapping legacy software screens for automation
- Preparing an agent to work in an interface
- Checking that the required elements are on screen
- Sizes
- under 1B (detector + captioning)
- Hardware
- from: Laptop
TextOllama2024
Nexusflow · USA
Fine-tuned Llama 3 and Qwen 2.5 models from Nexusflow. Athene-V2-Agent is specially trained for function calling and agent scenarios. Commercial use is prohibited.
- Research on agents and function calling
- Comparison with commercial models
- Experiments with a chat assistant
- Sizes
- 70B – 72B
- Hardware
- from: 1 GPU
