cua-s1-4b-0.2

LoRA adapters on the frozen Qwen/Qwen3.5-4B base for closed-option computer-use element/action decisions, plus agentic multi-step rollouts in live GUI environments. Two independently trained adapters, text/ and multimodal/.

Part of the Cua-S1 research family. Does not replace cua-s1-4b-0.1; both are separately published.

Results, methodology, and code: libs/cua-bench-s1, libs/cua-s1.

Usage

from cua_s1.four_b import FourBModel

model = FourBModel(
    base_model="Qwen/Qwen3.5-4B",
    lora_adapter_path="cua-ai/cua-s1-4b-0.2",  # picks text/ or multimodal/ by modality
    modality="text",
)

License

The adapter weights are licensed Apache-2.0. Qwen/Qwen3.5-4B's own weights and license govern the base model itself and are not redistributed here.