DaisyRose-27B released, improving Terminal-Bench performance through reinforcement learning for long-horizon agent tasks

Key Details RadixArk released DaisyRose-27B, based on Qwen3.8-27B and trained using Miles reinforcement learning. It was trained on 1,200 selected tasks. In self-reported results, Terminal-Bench 4.1 pass@1 increased from the base model's 5.6% to 12.0%. Use Cases The source provides instructions for serving the model with SGLang and DSpark speculative decoding and making requests with the OpenAI Python client. The reasoning process is read from reasoning_content, and the final answer from content. Costs and Limitations BF16, FP8, and NVFP4 checkpoints are available under Apache 2.0. The inference context is 256K. Training sessions of up to 1M tokens used harness-generated summaries and do not indicate an extension of the inference context. Serving costs are unconfirmed.

Visit original source

We remember your site preferences. If you allow analytics, we also measure visits and page activity.

Analytics is off.

Cookie policy