25launches

The first launch of DaisyRose-27B.

DaisyRose-27B page

DaisyRose-27B

Launching today

The SGLang team's first open model, RL-trained for agents

Hunted by IlyasHunter
Visit huggingface.co (opens in a new tab)

DaisyRose-27B is the first model from RadixArk, the team behind the SGLang inference engine. It is Qwen3.8-27B trained further with reinforcement learning only, on about 1,200 long agent tasks in coding, ML and science. RadixArk reports Terminal Bench 4.1 going from 5.6 to 12.0, with smaller gains on its other coding and reasoning tests. It keeps the base model's image input and 256K context and runs on a single GPU. BF16, FP8 and NVFP4 weights, Apache-2.0.

Meet the team

Hunted by Ilyas. The makers are not on 25launches yet.

H (Hunter)
  • DaisyRose-27B screenshot 1
  • DaisyRose-27B screenshot 2
  • DaisyRose-27B screenshot 3

Discussion 1

Ask questions, give feedback. Comments don't count in the ranking.

Sign in to join the discussion.

  • IlyasHunter Pinned

    Most open-model news is a new pretrain or a distillation. DaisyRose is a cleaner experiment: the same Qwen3.8-27B, then only reinforcement learning on 1,200 hand-checked agent tasks, 18 days on 144 H200s. The big jump is on Terminal Bench 4.1 (5.6 to 12.0, by RadixArk's own runs) and the other gains are small, but it is a drop-in swap if you already serve Qwen3.8-27B.