UI-TARS
https://github.com/bytedance/UI-TARS
ByteDance Seed's open-source GUI-agent foundation model — released Jan 2025; trained end-to-end on screenshot → action pairs. Comes with UI-TARS-Desktop (Electron app) that runs the model against the user's screen. SOTA on OSWorld at release; weights on HuggingFace.
Active · Verified 2026-10-31 · fresh as of 2026-10-31 · 49 cited facts, 3 with a third-party source · 2 recorded changes · Suggest a correction
At a glance
- Tier
- T2
- Section
- Computer-use & desktop agents
- Latest release
- UI-TARS-desktop / Agent TARS v0.3.1 (2026-10-09); UI-TARS-2 announced 2025-09-04; latest open weights UI-TARS-1.5-7B (2025-04-16) repo fix
- License
- Apache-2.0 repo fix
- GitHub
- bytedance/UI-TARS-desktop — ~39.5k stars, last commit 2026-10 (v0.3.1, 2026-10-09); model repo bytedance/UI-TARS — ~11.7k stars, last commit 2025-09 repo fix
- Pricing
- OSS weights free repo fix
- Funding
- ByteDance internal (Seed team) vendor fix
- Freshness
- 2026-10-31 fix
- Layer
- agent-infra repo fix
- Licensing
- open-source repo fix
- Deployment locus
- computer-use repo fix
- Developer activity
- OpenRank 0.66 (12-mo avg 1.93) · 4 participants — 2026-09 third-party fix
When to use
Optimised for: End-to-end GUI-agent foundation model — screenshot → action
Anti-fit: Research-grade — no SLA / hosted endpoint from ByteDance; 7B baseline is small
Pros & cons
Pros
First major open-weights GUI agent FM; SOTA at release; Apache 2.0; full training recipes.
Cons
ByteDance origin = some enterprise data-residency concerns; no hosted endpoint; 7B baseline.
Claims & capabilities
arXiv 2501.12326 (Jan 2025); OSWorld 24.6% beating Claude 3.5 Sonnet (22.0%); weights on HF
Technical surface
- API surface
- Python / VLLM; desktop app for end users
- Backend storage
- Caller-side; no built-in storage
- Deployment
- Self-hosted only — open weights (Hugging Face) + UI-TARS-desktop / Agent TARS app
- Embedding model
- not applicable — not a memory product
- Multi-tenancy
- Single-process per model instance
- MCP
- yes — via Agent TARS (UI-TARS-desktop): MCP-based kernel that mounts MCP servers
- A2A
- searched not found
- OpenTelemetry
- searched not found
Alternatives to UI-TARS → live options ranked by shared design, licence, hosting and MCP support
Similar systems
Other computer-use & desktop agents in the catalog, ranked by inbound references.
- Browser Use T1
Python library that turns any LLM into a browser agent — extracts DOM into a structured representation the LLM can act on (rather than pure screenshot). Browser Use Inc. raised $17M seed Q1 2025 (Felicis); cloud product at cloud.browser-use.com. Project is one of the most-starred browser-agent OSS repos.
- Stagehand T1
Open-source browser-automation framework by Browserbase (TypeScript core plus Python and Go SDKs). Mixes Playwright-style code with natural-language act/extract/observe; pairs with Browserbase Cloud.
- Hyperbrowser T2
Cloud-hosted headless browser API tuned for AI agents — captcha solving, stealth mode, session recordings. Y Combinator S24; pitched as a faster / cheaper Browserbase alternative. Used as substrate by agents that need browser execution but don't want to run their own Chromium fleet.
- Anthropic Computer Use T1
Anthropic's API-surface capability that lets Claude control a virtual desktop — see screenshots, type, click. Launched 2024-10 as a public beta. Not a finished IDE itself; it's the substrate that other harnesses (and Anthropic's own Claude apps) use to operate a computer.
- Skyvern T2
Open-source autonomous browser agent that uses computer vision + LLMs to navigate websites. Workflow-based: agents follow recorded plans that adapt to layout changes. Skyvern Inc. raised $4.5M seed (Sep-2024, Felicis lead). Self-hostable; cloud version on skyvern.com.
- Bytebot T3
Open-source desktop-agent runtime that runs an LLM-controlled Linux desktop inside Docker — uses xdotool / X11 capture for control. Pitched as the 'self-hosted Operator' alternative; agent sees a real desktop screen, not just a browser.
Sources (48 cited facts, 3 third-party)
Each fact links to the page it came from; the badge says who published it (how to read sources).
- What it is OSS GUI-agent foundation model + desktop app (ByteDance) repo
- What it does & what's distinct ByteDance Seed's open-source GUI-agent foundation model — released Jan 2025; trained end… repo
- Claims / benefits arXiv 2501.12326 (Jan 2025); OSWorld 24.6% beating Claude 3.5 Sonnet (22.0%); weights on… paper
- Modalities screen images + keyboard/mouse actions repo
- Created 2025-01 paper
- Latest release UI-TARS-desktop / Agent TARS v0.3.1 (2026-10-09); UI-TARS-2 announced 2025-09-04; latest… repo
- License Apache-2.0 repo
- GitHub bytedance/UI-TARS-desktop — ~39.5k stars, last commit 2026-10 (v0.3.1, 2026-10-09); mode… repo
- Mindshare Front-page HN; benchmarked SOTA on OSWorld at release paper
- Citations Cited in subsequent computer-use papers (2025) paper
- Funding ByteDance internal (Seed team) vendor
- Customers / scale OSS users; research community repo
- Pricing OSS weights free repo
- Data handling Self-host model; user controls data repo
- Deployment Self-hosted only — open weights (Hugging Face) + UI-TARS-desktop / Agent TARS app repo
- HQ China (ByteDance, Beijing) + Seattle research vendor
- Founders / pedigree ByteDance Seed (Doubao team) third-party
- Performance OSWorld 24.6%; AndroidWorld 46.6%; ScreenSpot-V2 91.6% (Jan 2025 paper) paper
- Reproducibility Weights + training recipes in repo repo
- Code/weights release OSS Apache 2.0 — github.com/bytedance/UI-TARS repo
- API surface Python / VLLM; desktop app for end users repo
- Backend storage Caller-side; no built-in storage repo
- Multi-tenancy Single-process per model instance repo
- Versioning Model versions: 1.0 / 1.5 / 2.0 on HF repo
- Namespace primitives per-model instance repo
- MCP support yes — via Agent TARS (UI-TARS-desktop): MCP-based kernel that mounts MCP servers repo
- Import / export Weights on HuggingFace registry
- Validated verticals horizontal GUI control; benchmarks span web + desktop + mobile paper
- Time to running <1day (download weights + serve) repo
- Anti-fit Research-grade — no SLA / hosted endpoint from ByteDance; 7B baseline is small repo
- Optimised for End-to-end GUI-agent foundation model — screenshot → action paper
- Adjacent infrastructure VLLM; Electron desktop app; HuggingFace repo
- Pros First major open-weights GUI agent FM; SOTA at release; Apache 2.0; full training recipes. paper
- Cons ByteDance origin = some enterprise data-residency concerns; no hosted endpoint; 7B basel… repo
- Project & sources github.com/bytedance/UI-TARS; arxiv.org/abs/2501.12326 repo
- commit-trajectory [{"ym":"2025-01","cum":52},{"ym":"2025-02","cum":56},{"ym":"2025-04","cum":85},{"ym":"20… repo
- cost-input-usd-per-mtok 0 repo
- cost-output-usd-per-mtok 0 repo
- cost-tier free repo
- cost-pricing-model free-tier repo
- cost-last-verified 2026-06-29 repo
- capability-benchmark-sources osworld: 53.1% (UI-TARS-2-2509, 100 steps, OSWorld-Verified leaderboard entry 2025-10-14… third-party
- architecture-deployment-locus computer-use repo
- licensing-bucket open-source repo
- positioning-platform-vs-product platform repo
- taxonomy-three-layer-rollup agent-infra repo
- developer-activity-opendigger {"activity":4.96,"as_of":"2026-09","openrank_12mo":1.93,"openrank_current":0.66,"partici… third-party
- sdk-languages python repo
What changed
Changes to this record's lifecycle, licence, release, pricing and hosting facts, from the catalog's git history. Subscribe (Atom)
- 2026-10-31
- Latest release “UI-TARS-2 announced (2025-09-04); latest open weights UI-TARS-1.5-7B (2025-04-16)” → “UI-TARS-desktop / Agent TARS v0.3.1 (2026-10-09); UI-TARS-2 announced 2025-09-04; latest open weights UI-TARS-1.5-7B (2025-04-16)” source
- 2026-10-26
- License “Apache 2.0” → “Apache-2.0” source
- Latest release “UI-TARS-1.5 (Apr 2025)” → “UI-TARS-2 announced (2025-09-04); latest open weights UI-TARS-1.5-7B (2025-04-16)” source
- Deployment “Self-host (HuggingFace weights) + UI-TARS-Desktop Electron app” → “Self-hosted only — open weights (Hugging Face) + UI-TARS-desktop / Agent TARS app” source
- MCP support first recorded as “yes — via Agent TARS (UI-TARS-desktop): MCP-based kernel that mounts MCP servers” source