Key takeaways

  • RadixArk released DaisyRose-27B on November 18, 2026, under the Apache 2.0 license.
  • The company says it solves 12.0% of Terminal Bench 4.1 tasks on the first try, up from 5.6% for the Qwen3.8-27B base model.
  • Training took 18 days on 144 H200 GPUs, about 63,000 GPU-hours.

RadixArk, which lists the SGLang inference engine and the Miles training framework among its open-source projects, released DaisyRose-27B on November 18, 2026, its first model. It is Alibaba's Qwen3.8-27B with extra reinforcement learning on long agent tasks, and the weights are on Hugging Face under the Apache 2.0 license, in BF16, FP8 and NVFP4 versions.

The biggest gain is on hard terminal work. RadixArk says DaisyRose-27B solves 12.0% of Terminal Bench 4.1 tasks on the first try, up from 5.6% for the base model. On other tests the change is small: 75.5 against 73.0 on Terminal Bench 2.1, 46.2 against 42.2 on DeepSWE v1.1, and 34.3 against 33.9 on Humanity's Last Exam (text only). All of these are the company's own numbers, run with its own harnesses; no independent scores are out yet.

The training recipe is the more interesting part for teams that post-train their own models. RadixArk trained on about 1,200 curated tasks, each up to 1.5 hours long with more than 100 tool calls, chosen to target mistakes the base model makes. The run took 18 days on 144 Nvidia H200 GPUs, about 63,000 GPU-hours, using Miles. RadixArk says its training backend cut the cost by 57% compared with an open-source Megatron-based trainer, and it ran about 944,000 sandboxes for the agent tasks.

For developers, it is a dense 27-billion-parameter model aimed at long coding and terminal sessions, small enough to run on hardware a team already owns. Whether the doubled Terminal Bench 4.1 score holds up outside RadixArk's harness is the thing to check first. Our article on Tencent's Hy4 flash covers another open-weight release from the same week.

Sources

  1. DaisyRose-27BRadixArk
  2. RadixArk/DaisyRose-27B model cardHugging Face

Add ClaudeGPT Si as a preferred source on Google