UI-TARS Desktop: ByteDance's Open-Source Computer Use (31k Stars)

追加: 2026-05-16

[00:00:00]Bite Dance's open- source computer use is at 31K Gate Hub stars. UITAR's desktop runs a vision language model behind a native Electron app that controls your computer or browser locally or remotely. Same idea as anthropic computer use open weights underneath. UIR's desktop is Biteance's native GUI agent for your local computer. The Redmi puts it directly, a native GUI agent for your local computer driven by UITARS and seed 1.5VL 1.6 series models. Headline tagline, the open-source multimodal AI agent stack, connecting cutting edge AI models and agent infrastructure. The shape is the same as anthropic computer use or open AI operator. Natural language in mouse clicks and keyboard inputs out with the agent reading raw screenshots to figure out what's on screen. The difference, this one's open source, runs locally, and you bring your own model. Four control surfaces local computer local browser remote computer remote browser same agent same interface four targets 31.3k gate hub stars Apache 2.0 codebase 89% TypeScript native electron app crossplatform on Windows and Mac OS latest tagged release 0.3.0 zero from November. The category here matters because UITAR's desktop sits in the same product space as a handful of recently shipped commercial agents. Anthropic computer use lets Claude move your mouse. Open AI operator does the same with GPT. Open interpreter does it with whatever LLM you wire up. UITRS desktop is the open- source entry built around Bite Dance's purpose-built UIRS family of vision language models, but with the option to swap in Claude or any compatible model if that's the choice you'd rather make. The agents input is screenshots, full stop. It doesn't parse HTML, doesn't query DOM trees, doesn't use accessibility APIs. It reads pixels and decides what to click. That's both the limitation and the unlock.

[00:02:00]Limitation because it's slower and more compute-hungry than DOM scraping. Unlock because it works on absolutely anything that renders to a screen. Native apps, weird electron things. Lock down web UIs that defeat selectors. If a human can use it, the agent can use it. The four control surfaces are the actual feature matrix. Local computer. The agent runs on your machine and controls your machine. Mouse, keyboard, screen. Local browser. The agent runs on your machine and controls a browser instance which is the safer subset for most people who don't want autonomous code running with system level permissions. Remote computer. The agent runs on your machine but controls a different machine exposed through biteance's remote computer operator service introduced in v 0.2.0.

[00:02:44]Remote browser. Same idea for a hosted browser instance. The remote operators are free which is unusual. Most equivalent commercial offerings charge per session or per minute. Same natural language interface across all four.

[00:02:57]You'd pick local computer for personal automation tasks. Local browser for safer scripted browsing. Remote computer for sandbox dev or QA work. Remote browser for scraping or testing without exposing your local environment.

[00:03:11]Different tools for different threat models all in one app. Install is straightforward. npx at agent-tars CLI at latest pulls and runs the agent CLI without installing globally or npm install-g for the global install requires node 22 or newer then you launch the agent with a provider flag and a model flag and your API key two provider patterns the redmem documents bulk engine which is bite danc's hosted endpoint for the dubau and uitars family with the model flag pointing at one of the duba-1-5- thinking vision -provariants. That's the path if you want bite dances purpose-built GUI agent model or anthropic with the model flag pointing at claude-3-7- sonnet- latest. That's the path if you'd rather route through claude. Either way, the agent starts up, captures the screen, waits for instructions. From there, you describe what you want done.

[00:04:07]The agent reads the screen, decides the actions, fires the mouse and keyboard, observes the result, repeats side by side, the two provider configs. Bulk engine on top with the Dubau 1.5 thinking vision pro model purpose-built by bite dance for guey agent work. This is the model UI's desktop is optimized for anthropic on the bottom with claude 3.7 sonnet. Slightly different reasoning style, often better at multi-step tasks but slower per screenshot. Same flags, different provider, different cost profile. The agent itself doesn't change. You pick the model that fits the cost, latency, and accuracy trade-off for your workflow. The agent CLI also accepts other OpenAI compatible providers via the same pattern. So you can route through Open Router or your own enterprise gateway if that's how your org buys model access. On the model side, the UITARS family itself is worth understanding because it's open weights.

[00:05:00]UITARS comes in three sizes 2B, 7B and 72B parameters. The 7B model, UITARS 1.5 to 7B is on hugging face downloadable, runnable on a single consumer GPU. Bite dance trained these models specifically for GUI agent work, screen detection, action prediction, click coordinates rather than general vision language tasks adapted later for automation.

[00:05:25]According to bite danc's published benchmarks plus venture beats coverage the family beats GPT40 claude and gemini on more than 10 gui specific benchmarks whether those numbers hold in your workload depends on what you're automating but the trajectory is clear this is a category where purpose-built vision models outperform general ones and bite dance is shipping the most capable open weights option in that category desktop is the application the model is the engine both are first party from the same team by the numbers.

[00:05:57]31.3K GitHub stars on bite dance/UI-tars- desktop. Apache 2.0. No commercial restrictions on the code. Codebase 89% TypeScript packaged as a native Electron app for cross-platform desktop deployment. Latest tagged release v0.3.0 from November 4th, 2025. V0.2.0 added the free remote computer operator and remote browser operator. UI tars the model line three sizes 2B 7B 72B parameters open weights on hugging face for the smaller two performance claim per bite dance and venturebeat state-of-the-art on 10 plus UI benchmarks beats GPT40 claude Gemini built and maintained by biteance's seed team the companion repo bite dance/UI-tars holds the model weights and training details the line that captures the design philosophy is in the readme itself self the open- source multimodal AI agent stack connecting cuttingedge AI models and agent infrastructure. The whole bet is that the model layer and the agent layer are separate concerns and both should be open. Enthropic computer use is closed model plus closed agent. Open AI operator is closed model plus closed agent. Open interpreter is open agent plus bring your own model.

[00:07:15]UIR's desktop is open agent plus open model. The model layer is also from bite dance also Apache 2.0 Ow weightights also runnable locally. That's the strongest position in the open category.

[00:07:26]If you want to build something on top of GUI agent technology, automated testing, accessibility tooling, local only RPA, you can fork the entire stack, no vendor lockin at any layer. That's a meaningfully different deal than the commercial alternatives. And 31K stars say a lot of people have noticed. UIR's desktop open- source computer use 31K stars. your model behind it. Four targets to control. Start the repo.

[00:07:52]Install the CLI.