---
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B
datasets:
- OS-Copilot/OS-Shepherd-100K
language:
- en
tags:
- computer-use
- reward-model
- trajectory-evaluation
- multimodal
- mixture-of-experts
---

# OS-Shepherd-35B-A3B

OS-Shepherd-35B-A3B is an open multimodal reward model for judging computer-use agent trajectories. Given a task instruction, screenshots, and the agent's reasoning and actions, it determines whether the task was completed and returns a reasoned `SUCCESS` or `FAIL` verdict.

It is fine-tuned from [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) on [OS-Shepherd-100K](https://huggingface.co/datasets/OS-Copilot/OS-Shepherd-100K) using the same SFT and GRPO recipe as OS-Shepherd-9B. The RL stage focuses on reducing false-success judgments.

## Results

| Benchmark | Accuracy | Fail recall |
|---|---:|---:|
| OSReward | 85.6 | 86.2 |
| OSReward-Hard | 62.7 | 60.1 |

Results use the fixed judging protocol described in the [OSReward paper](https://arxiv.org/abs/2607.28609).

## Usage

Use the canonical prompt and trajectory format from the [OSReward repository](https://github.com/OS-Copilot/OSReward). A recent Transformers, vLLM, or SGLang version with Qwen3.5 multimodal support is required.

This model is intended for trajectory evaluation, data filtering, and reward-model research. It is not a computer-control policy and may still miss fine-grained visual failures, especially on hard cases.

## License

Apache License 2.0. See `LICENSE`.

## Citation

```bibtex
@article{sun2026osreward,
  title={OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models},
  author={Sun, Qiushi and others},
  journal={arXiv preprint arXiv:2607.28609},
  year={2026}
}
```