Physix Frontier · News Briefing Card (HuggingFace · Nov 19, 2026)

RadixArk releases DaisyRose-27B, doubling terminal task pass rate

KEY FACTS

  • RadixArk released its first model, DaisyRose-27B, trained with reinforcement learning on top of Qwen3.8-27B.
  • Training used 1200 tasks and took 18 days on 144 H200 GPUs.
  • pass@1 on Terminal-Bench 4.1 rose from 5.6% to 12.0%.
  • The model keeps the original architecture and tokenizer, and supports 256K context and image understanding.
  • The weights are released under the Apache 2.0 license, and the team officially recommends deploying inference with SGLang.

KEY DATA

5.6% → 12.0%Terminal-Bench 4.1 p
1200Training tasks
18 daysTraining duration
144 H200 GPUsTraining compute

PHYSIX OBSERVATION

A small team doubled a 27B model's terminal task pass rate using just over a thousand carefully curated tasks and an asynchronous RL backend, which shows that data quality and engineering efficiency in post-training matter more than piling on compute. With open weights and 256K context, it is a directly usable base for developers who want to build local agents, and its reproduction cost is worth tracking.

Source: HuggingFace report