DaisyRose-27B doubles Terminal-Bench 4.1 score with just 1,200 training tasks
DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

RadixArk trained DaisyRose-27B on Qwen3.8-27B using reinforcement learning on long-horizon agent tasks, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. The run used only 1,200 curated tasks over 18 days on 144 H200 GPUs. Gains also appeared across reasoning, instruction-following, and coding benchmarks. A custom training backend cut GPU time by 57%, and an in-house sandbox service ran roughly 944,000 sandboxes at about 80% lower cost than cloud providers. BF16, FP8, and NVFP4 checkpoints are released.
Each task targets a mistake the original model makes: we ran the original Qwen3.8-27B on every training task and carefully redesigned each one so that it brings out the mistakes we want the model to stop making.
- woah
The heavy reliance on eval datasets showcased here and also evidenced by the frontier lab's high payments for such data cast doubt on the claims that AI will become self-improving soon
- jacquesm
I always wonder what you lose when you start maxing out some specific benchmark. It is really hard to do this in a balanced way and I would love if everybody that did these kind of efforts would take the time to also list the before/afters of a number of other benchmarks that they did not use in their training.
- pkoiralap
Sabi-noema also looks like a close competitor in the same space. Built from qwen3.8 and made better.
https://huggingface.co/sabicap/sabi-noema
Also HF model here:
- xlayn
Snap, no gguf to download, will wait for one to appear to test, any improvement over an already good model is good news.
"18 days on 144 H200 GPUs" insane.. didn't thought so much hardware would be needed to train Qwen.
- mips_avatar
How did you guys set up the RL topology? Were you guys doing individual RL nodes (4xh200 rollout/4xh200 train) or did you have dedicated trainer and rollout nodes?