AI TrackerGet the app
← All news
ModelsOpen source

RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned for agent tasks that lifts Terminal-Bench 4.1 from 5.6% to 12.0%

RadixArk says DaisyRose-27B is its first model, built on Qwen3.8-27B and trained with reinforcement learning on long-horizon agent tasks, with sessions up to 1 million tokens. It says it used about 1,200 curated tasks, made with Eigent AI, over 18 days on 144 H200 GPUs, using its open-source Miles framework.

These are RadixArk's own results, run with its internal mini-swe-agent harness:

  • Terminal-Bench 4.1 pass@1: 12.0%, up from 5.6% for the original Qwen3.8-27B. RadixArk says it also scores higher than the original on every other benchmark it lists (Terminal Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, HLE text-only, IFBench and GPQA Diamond).
  • Training cost: RadixArk says its internal training backend reached 37% model FLOPs utilization against 16% for the open-source Megatron-based trainer, so the same training needs 57% less GPU time.
  • Sandboxes: it says it ran about 944,000 sandboxes for the project on its own service built on the open-source AgentENV project, at about 80% lower cost than popular cloud sandbox providers.

BF16, FP8 and NVFP4 checkpoints are on Hugging Face. RadixArk says it is offering the same approach to teams that want specialized models.

Get news like this as a push.

AI Tracker alerts you when something changes for the tools you use.

Get the app