Models

← All models

OpenCUA-Qwen2-7B

XLANG Lab · 2025-08-13

OpenCUA-Qwen2-7B is a paper-reported computer-use model obtained by Stage-2-only fine-tuning of Qwen2-VL-7B-Instruct for 30 billion tokens. The paper says the Qwen base was adapted to use 1D RoPE plus Kimi-VL-A3B's tokenizer and chat template and trained with action-only, thought-plus-action, and observation-plus-thought-plus-action formats; no standalone checkpoint for this variant is currently publicly linked, so Atlas retains it as a benchmark entry distinct from the released Qwen2.5-VL-based OpenCUA-7B.

Benchmark scores