RadixArk has published DaisyRose-27B, a 27-billion-parameter model built on Qwen3.8-27B and further trained with reinforcement learning using Miles. According to RadixArk, the team focused on long-horizon agent tasks, running sessions as long as 1 million tokens and training on about 1,200 curated tasks over 18 days on 144 H200 GPUs.
Reported benchmark gains and model properties
RadixArk states DaisyRose-27B raises Terminal-Bench 4.1 pass@1 from 5.6% (Qwen3.8-27B) to 12.0%. The release notes also show improvements across other benchmarks: Terminal Bench 2.1 (73.0 to 75.5), DeepSWE v1.1 (42.2 to 46.2), SWE Bench Pro v2 (91.3 to 94.7), HLE text-only (33.9 to 34.3), IFBench (79.5 to 82.4) and GPQA Diamond (89.2 to 90.7). RadixArk reports scores use the higher of BF16 and FP8 runs and that the two formats differ by less than one point where both were evaluated.
The model retains Qwen3.8-27B’s architecture, tokenizer, chat template, image understanding and a 256K-token context window. Checkpoints are provided under an Apache 2.0 license in BF16, FP8 and NVFP4 formats, RadixArk says.
Training pipeline and infrastructure
RadixArk attributes the release to three elements: curated tasks developed with Eigent AI, an asynchronous training setup using Miles, and a large-scale sandbox service based on AgentENV. The training run used 128 H200 GPUs for agent inference while 16 GPUs ran the trainer; updated weights were propagated to the inference workers after each step, RadixArk explains.
RadixArk reports a model-FLOPs utilization (MFU) of 37% on its backend versus 16% for a comparable open-source trainer, which it says corresponds to 57% less GPU time. The team also reports about 2× faster end-to-end throughput on the release run (roughly 16,800 tokens/sec across 144 GPUs versus ~8,300 tokens/sec for the open-source trainer) and cites larger efficiency gains for mixture-of-experts variants.
Between mid-September and mid-November the sandbox service ran approximately 944,000 sandboxes with up to 1,566 concurrent instances; RadixArk reports sandbox creation success around 99% overall and above 99.95% since late October. The project’s sandbox cloud costs are reported as about 80% lower than popular cloud sandbox providers.
RadixArk recommends serving DaisyRose-27B with SGLang and DSpark speculative decoding. The model card and quickstart note that a local server started with the model name DaisyRose-27B on port 30000 can be queried with an OpenAI-compatible client; RadixArk states the model returns reasoning in reasoning_content and the final answer in content. Checkpoints and serving guidance are available for download, and RadixArk invites users to share results in the model’s community tab.
Original source: Hugging Face Blog