RadixArk Releases DaisyRose-27B Model
Reported by Hugging Face1 source ↓
RadixArk has released DaisyRose-27B, its first model, trained with reinforcement learning on long-horizon agent tasks. The model is based on Qwen3.8-27B and is available to download.
Key points
- DaisyRose-27B is RadixArk’s first model and is based on Qwen3.8-27B
- RadixArk reports Terminal-Bench 4.1 pass@1 improved from 5.6% to 12.0%
- Training used 1,200 selected tasks and took 18 days on 144 H200 GPUs
- The model retains a 256K-token context window
- RadixArk recommends serving it with SGLang and DSpark speculative decoding
Details
Hugging Face reports that RadixArk trained DaisyRose-27B with reinforcement learning using Miles, its open-source framework. Training used long-horizon agent tasks, with some sessions reaching 1 million tokens; harness summarization was used to span multiple trajectories. The model retains the architecture, tokenizer, chat template and image understanding of Qwen3.8-27B.
Its largest reported gain was on Terminal-Bench 4.1, where pass@1 rose from 5.6% to 12.0%. The company says scores use the higher of BF16 and FP8 runs, which differed by less than one point where both were tested.
The training run lasted 18 days and used 144 H200 GPUs, according to RadixArk. The company reports 63,000 GPU-hours in total and says its internal backend achieved 37% model FLOPs utilization, compared with 16% for the open-source trainer on comparable runs. RadixArk also reports that training throughput, including agent task time, was about 16,800 tokens per second, versus about 8,300 for the open-source trainer.
RadixArk says the model is available to download and recommends SGLang and DSpark speculative decoding for serving it. The company also reports that its sandbox service ran about 944,000 sandboxes for the project between mid-September and mid-November, with up to 1,566 running at once.
Written from coverage as of
Images
- Hugging Face
- Hugging Face
- Hugging Face