UI-TARS Desktop overview: models and control modes

UI-TARS Desktop is presented as ByteDance's open source AI agent for controlling a computer or browser through screenshots, mouse clicks and keyboard input. The overview compares its approach with Anthropic Computer Use, OpenAI Operator and Open Interpreter. The speaker describes an Electron app for Windows and macOS, with Apache 2.0 code, and identifies v0.3.0, dated November 4, 2025, as the latest tagged release covered.

The four control modes target a local computer, local browser, remote computer or remote browser. The speaker says the remote operator services arrived in v0.2.0. Screenshot-based control avoids reliance on HTML, DOM trees or accessibility APIs, but the speaker describes it as slower and more compute hungry than DOM scraping.

The video also introduces the separate Agent TARS CLI, using npx @agent-tars/cli@latest with Node 22 or newer. Provider and model flags select a hosted VolcEngine endpoint or Anthropic's Claude 3.7 Sonnet, with an API key. The walkthrough also notes support for other OpenAI-compatible providers. These CLI examples use hosted model APIs, so local agent execution does not keep the model requests on the machine.

For local AI, the video discusses UI-TARS models in 2B, 7B and 72B sizes. The speaker says UI-TARS-1.5-7B is downloadable from Hugging Face and can run on a single consumer GPU, without specifying VRAM requirements. Benchmark advantages are attributed to ByteDance and VentureBeat; the speaker cautions that results depend on the workload.