RadixArk releases DaisyRose-27B, trained from Qwen3.8-27B
Original title: DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1
RadixArk says 1,200 reinforcement-learning tasks raised Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with its Miles reinforcement-learning framework. It says training used about 1,200 curated long-horizon tasks over 18 days on 144 H200 GPUs, with sessions up to 1 million tokens. BF16, FP8 and NVFP4 checkpoints are on Hugging Face. Benchmark gains, including other reasoning and coding tests, are the company’s own figures.
Unconfirmed: Benchmark and efficiency figures are RadixArk’s own measurements, not an independent evaluation.
- Training tasks were built with Eigent AI and capped at about 1,200.
- The release run used 63,000 GPU-hours on 144 H200s, RadixArk says.
- Sandboxes ran on AgentENV; about 944,000 were created for the project.
- Miles is the open-source trainer RadixArk says others can use.
- Products
- Miles
- AI models
- DaisyRose-27BQwen3.8-27B
From the original article
5 min read at radixark.com
Today we are releasing DaisyRose-27B , RadixArk’s first model, built on Qwen3. 8-27B and trained with reinforcement learning using Miles . We trained it on long-horizon agent tasks with sessions up to 1M tokens, with just 1,200 curated tasks in 18 days on 144 H200 GPUs.
By de_sty · radixark.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.