RadixArk on Nov. 18 released DaisyRose-27B, a Qwen3.8-27B fine-tune for agent tasks trained in 18 days on 144 H200 GPUs, and said its training backend needed 57% less GPU time.
The efficiency claim is the core of the release. If it holds, teams running reinforcement learning on long agent sessions would need fewer GPU hours for the same run. Every figure comes from RadixArk's own announcement on Hugging Face1, which includes no independent test results. The model's headline benchmark score is 12.0%.
RadixArk describes DaisyRose-27B as its first model. It was trained with Miles, the company's open-source reinforcement learning framework, and the post presents the release as a demonstration of post-training efficiency. Agent tasks ran in sandboxes on AgentENV2, an open-source project that, according to its README, also powers agentic RL training for Kimi K3. RadixArk said it contributes fixes to AgentENV.
Benchmark results
DaisyRose-27B raised Terminal-Bench 4.1 pass@1 to 12.0% from 5.6% for the base model, RadixArk said. The post also says the model scores higher than Qwen3.8-27B on every benchmark in its results table, and describes gains on "a variety of other reasoning, instruction following, and coding benchmarks." The post's text does not name those benchmarks or give their scores.
Scores use the higher of the BF16 and FP8 runs for each benchmark, which RadixArk said differ by less than one point wherever both were evaluated. The post says the model card includes comparison sources and evaluation settings; this article does not draw on the model card.
The model keeps Qwen3.8-27B's architecture, tokenizer, chat template, image understanding and 256,000-token context window. Checkpoints are available under Apache 2.0, and RadixArk recommends serving them with SGLang and DSpark speculative decoding.
Efficiency claims
RadixArk said its backend reached 37% model FLOPs utilization versus 16% for a Megatron-based open-source trainer on comparable runs, which it says means 57% less GPU time. The post does not describe the scale of those comparison runs. It also claims larger gains on mixture-of-experts models: 6.2 times on Qwen3.6-35B-A3B and more than 10 times on the roughly 300B-parameter GLM-5.3-Flash.
Counting time the agent spends on tasks, the release run processed about 16,800 tokens per second across its 144 GPUs, against about 8,300 for the open-source trainer on the same GPUs, RadixArk said. It said the run consumed 63,000 GPU-hours and that speed and efficiency held steady from first step to last.
Training ran in fully asynchronous mode. According to RadixArk, 128 H200s handled inference on agent tasks while 16 ran the trainer, and the updated model went out to the inference GPUs after every step. Each step learned from a handful of very long sessions, some up to 1 million tokens, folded into multiple trajectories through what the company calls harness summarization.
On the sandbox service, RadixArk said cloud infrastructure costs were about 80% lower than running the same sandboxes on "popular cloud sandbox providers," which the post does not name. It said 99% of sandbox creations succeeded over nine weeks and that the rate has exceeded 99.95% since late October.
Training data
RadixArk built the tasks with Eigent AI, starting from thousands of candidates drawn from Eigent's pool of tasks in machine learning, software engineering, science and other fields. About 1,200 made the final set. The company said it ran Qwen3.8-27B on each task, redesigned tasks around the model's mistakes and tuned difficulty until solve rates fell within "a useful learning range."
RadixArk also said it reviewed every task alongside the model's attempts, and that tasks which could be passed without solving the intended problem, or that had a design flaw, were fixed or removed. Each attempt could run for up to 1.5 hours in a sandbox and involve more than 100 tool calls and more than 256,000 output tokens, RadixArk said.