DaisyRose-27B
RadixArk released DaisyRose-27B, its first model, a Qwen3.8-27B fine-tune trained with reinforcement learning that raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.
RadixArk trained the model with its open-source Miles framework on about 1,200 curated long-horizon agent tasks, with sessions up to 1M tokens, in 18 days on 144 H200 GPUs. It says its internal training backend reached 37% MFU against 16% for the open-source trainer, which cuts GPU time by 57%. BF16, FP8 and NVFP4 checkpoints are available on Hugging Face.
- Date
- Wednesday 18 November 2026
- Lab
- RadixArk
- Kind
- open-weights
- Access
- open weights
- Price
- $0.15 per M input, $1.875 per M output on OpenRouter (DeepInfra, listed 25% off $0.20 / $2.50)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Terminal-Bench 4.1 pass@1 | 12.0% (base 5.6%) averaged over 8 runs per task, internal mini-swe-agent harness with summarization | company |
| Training MFU | 37% vs 16% for open-source trainer 57% less GPU time for the same training | company |
| Training compute | 63,000 GPU-hours on 144 H200s | company |
| Sandbox cost | about 80% lower than popular cloud sandbox providers about 944,000 sandboxes run | company |
Built on Qwen3.8-27B with training tasks made with Eigent AI. The license was not stated on the page.
Sources
This record was checked against its sources on 19 November 2026. How we check