UI-TARS 1.5 community proposal: local inference transport with explicit action-parser boundaries

#20
by Raullen - opened

Proposing a model-specific community integration note for UI-TARS-1.5 on an Apple Silicon local inference server.

Rapid-MLX has dedicated UI-TARS reasoning/action parsers. The linked tests cover recognized click/drag/hotkey shapes and streaming action holdback. This is more specific than a generic OpenAI-compatible text endpoint, but parser tests do not establish a working computer-use agent.

A useful example should show screenshot preprocessing, the coordinate space of returned actions and the client's execution boundary. An end-to-end screenshot/action check is required before claiming task completion. The initial scope would be one 1.5-7B quantization; DPO/SFT, 72B and other bit widths need separate evidence. The proposed documentation would help clients avoid treating a parsable action as a verified screen interaction.

Concrete catalog entry: ui-tars-1.5-7b-4bit → mlx-community/UI-TARS-1.5-7B-4bit@cca0e360e25322cdf480cf841e301b6d55bce9e0. This is a conversion/distribution artifact; it is not presented as an official upstream release.

Source references (fixed Rapid commit): catalog, action parser tests, stream holdback tests.

Verification status: this proposal is based on reading the catalog, implementation and the linked documentation/tests. No fresh model inference or test-suite run was performed for this submission; source tests are not an end-to-end qualification. It requests a community documentation scope, not an official endorsement or a performance listing.

Disclosure: I maintain Rapid-MLX. Prepared with AI assistance under my direction. Anonymous usage-metadata telemetry is enabled by default; DO_NOT_TRACK=1 disables it. The original model and conversion terms continue to apply.

Sign up or log in to comment