DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
Blog post from Hugging Face
RadixArk has released DaisyRose-27B, an Apache 2.0 model based on Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks. Trained over 18 days using 144 H200 GPUs and about 1,200 curated tasks, it raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than its base model across all reported benchmarks, including coding, reasoning, and instruction-following tests. DaisyRose retains the base model’s architecture, tokenizer, image understanding, and 256K-token context window, while training sessions could span up to one million tokens through trajectory summarization. RadixArk attributes the training run to carefully designed multi-step tasks, its asynchronous Miles training framework, and a scalable sandbox service; it reports lower GPU use and faster end-to-end training than comparable runs with an open-source trainer. The model’s BF16, FP8, and NVFP4 checkpoints are available, with serving guidance for SGLang and DSpark speculative decoding.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Reinforcement learning | 2 | 99 | 55 | 28 | -39% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.