DaisyRose-27B: Qwen Fine-Tuning Trained Exclusively via Reinforcement Learning

RadixArk's first model, uploaded on 11 18: Based on Qwen3.8-27B, with no supervised fine-tuning or distillation, using only reinforcement learning. The training tasks consist of approximately 1,200 long-chain agent tasks selected in collaboration with Eigent AI (each with a maximum sandbox duration of 1.5 hours and more than 100 tool calls). The maximum length of a single session is 1M tokens, and training was conducted for 18 days on 144 H200 GPUs. It retains the base model's architecture, tokenizer, chat template, vision capabilities, and 256K context window, allowing for a direct replacement of Qwen3.8-27B. The weights are licensed under Apache-2.0, with FP8 and NVFP4 versions also available.

Original title: RadixArk/DaisyRose-27B

RadixArk's first model, uploaded on 11 18: Based on Qwen3.8-27B, with no supervised fine-tuning or distillation, using only reinforcement learning. The training tasks consist of approximately 1,200 long-chain agent tasks selected in collaboration with Eigent AI (each with a maximum sandbox duration of 1.5 hours and more than 100 tool calls). The maximum length of a single session is 1M tokens, and training was conducted for 18 days on 144 H200 GPUs. It retains the base model's architecture, tokenizer, chat template, vision capabilities, and 256K context window, allowing for a direct replacement of Qwen3.8-27B. The weights are licensed under Apache-2.0, with FP8 and NVFP4 versions also available.

Why it matters The architecture, tokenizer, and chat template are identical to the base model, enabling direct replacement. Officially, most programming and terminal benchmarks show improvements; it can run on a single GPU and is under an Apache license.

Upvotes 42 · Likes 44 · Downloads 0

Reasons

  • Officially, Terminal Bench 4.1 scores rose from 5.6 to 12.0; Terminal Bench 2.1 73.0 → 75.5, DeepSWE v1.1 42.2 → 46.2, IFBench 79.5 → 82.4, and GPQA Diamond 89.2 → 90.7.
  • In the same table, significantly larger models like MiMo V2.6 Flash and Inkling 2 Small still score higher on most programming benchmarks; scores for other models are taken from official announcements or third-party leaderboards.
  • The reinforcement learning framework Miles is open-source: 128 GPUs for agent execution and 16 GPUs for training, totaling approximately 63,000 GPU hours.

Usage

  • Officially recommended: SGLang (lmsysorg/sglang:latest) with DSpark speculative decoding: --speculative-algorithm DSPARK --speculative-draft-model-path RadixArk/Qwen3.8-27B-DSpark, plus --reasoning-parser qwen3 --tool-call-parser qwen3_coder; the SGLang cookbook generates launch commands based on GPU and precision.
  • Weights are approximately 52 GB (BF16) / 29 GB (FP8) / 23 GB (NVFP4, requires Blackwell GPUs); officially, it can run on a single GPU such as RTX 5090, RTX PRO 6000, or DGX Spark.
  • For thinking mode, use temperature 1.0, top_p 0.95, top_k 20, and min_p 0; for difficult problems, do not set max_tokens, and ensure reasoning_content is passed back in multi-turn conversations.

Risks

  • Thinking mode is enabled by default; officially, difficult problems may require tens of thousands of tokens, resulting in high latency and cost.
  • Scores are self-reported by the official team, while comparison model figures come from official announcements and third-party leaderboards, with inconsistent evaluation frameworks; the absolute score for Terminal Bench 4.1 remains only 12.0.
  • The official documentation does not specify which precision to use or how long the context window can be on GPUs like the RTX 5090 with 32 GB of VRAM. This entry has not been tested.

Open

Hugging Face Trending

Explore the latest AI Radar updates