DaisyRose-27B released
The story so far
RadixArk released the DaisyRose-27B model based on Qwen3.8-27B, trained through reinforcement learning (Miles) and long-term proxy tasks in 18 days using 144 H200 GPUs, improving the pass@1 of Terminal-Bench 4.1 from 5.6% to 12.0% and achieving gains on other reasoning, instruction following, and coding benchmarks. The model's checkpoint is provided under the Apache 2.0 license, recommended for speculative decoding using SGLang and DSpark.
AI-written from the reports · Updated 1 hour ago
Coverage timeline
Follow the coverage to see every side of the story.
- Hugging Face BlogDaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
RadixArk has released the DaisyRose-27B model based on Qwen3.8-27B, trained through reinforcement learning (Miles) and long-term proxy tasks (up to 1M tokens), completing fine-tuning in 18 days using 144 H200 GPUs, which increased the pass@1 of Terminal-Bench 4.1 from 5.6% to 12.0% and achieved improvements on other reasoning, instruction following, and coding benchmarks.
Heat over time
Not enough data yet to draw a trend.