RadixArk Releases DaisyRose-27B, a Reinforcement Learning Model Based on Qwen3.8-27B

·2026.11.20 19:30

Key point

RadixArk released DaisyRose-27B, trained on Qwen3.8-27B using reinforcement learning, under the Apache 2.0 license.

1 / 4

Details

  • Released DaisyRose-27B, trained on Qwen3.8-27B using only reinforcement learning without SFT via the open-source RL framework Miles
  • 1,200 long-term agent tasks, 144 H200 GPUs, 18-day training, and 256K context support
  • Recorded a score of 12.0 on Terminal Bench 4.1, but it is difficult to view this as a pure training effect due to differences in measurement conditions
  • Recommended combination of SGLang and DSpark speculative decoding; using a draft model yields up to 3.16x speed improvement per H200 (measured on base model)

Training and Benchmarks

  • Uses the same architecture and tokenizer as Qwen3.8-27B, ensuring compatibility with existing inference environments
  • Provides weights in BF16 52GB, FP8 29GB, and NVFP4 23GB
  • Achieved scores of 82.4 on IFBench, 94.7 on SWE Bench Pro v2, and 90.7 on GPQA Diamond

Infrastructure and Serving

  • Uses Miles in fully asynchronous mode, with 128 H200 GPUs dedicated to inference and 16 to training
  • The draft model Qwen3.8-27B-DSpark is trained based on Qwen3.8-27B and differs from DaisyRose-27B
  • The draft model license is listed as 'other', requiring separate verification for commercial use

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.