DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
- Source
- huggingface.co
- Date

Context
RadixArk, the company behind the open-source reinforcement learning framework Miles, has released DaisyRose-27B, its first model. It is Qwen's Qwen3.8-27B with extra training on about 1,200 long tasks, and the weights are free to use under the Apache 2.0 license. The release post is mainly a claim about how cheaply that training ran. The headline result, Terminal Bench 4.1 rising from 5.6% to 12.0%, is weaker evidence than it looks, because the two numbers were not measured the same way.
Reinforcement learning on agent tasks works by letting the model attempt a job in a , a fresh copy of the task's environment where it can run commands and edit files, then grading the result and nudging the model toward what worked. RadixArk says each task was built with Eigent AI around a mistake the base model makes, tuned so the model sometimes but not always succeeds, and allowed up to 1.5 hours and more than 100 tool calls per attempt. Training ran asynchronously: 128 H200 GPUs ran the agent on tasks while 16 more learned from the finished attempts, and updated weights went back to the agent machines after every step. The run took 18 days on 144 GPUs, about 63,000 GPU-hours.
The efficiency claim rests on one measurement. RadixArk reports that its training backend kept 37% of each GPU's peak compute doing useful training work, against 16% for an open-source trainer built on Megatron in comparable runs, so the same training needs 57% less GPU time, and that end-to-end throughput was about twice as high. Those are the company's own measurements, and the backend that produced them is described as internal, not as part of the open-source Miles release. It also says its sandbox service cost about 80% less than popular cloud sandbox providers.
The benchmark table shows DaisyRose ahead of the base model on all seven tests, by less than five points on every test except Terminal Bench 4.1. The model card explains where the numbers come from. DaisyRose's 12.0 is RadixArk's average over eight runs per task using its own internal agent setup with summarization turned on. The base model's 5.6 is a Terminal Bench 4.0 score that Artificial Analysis measured, used because Qwen3.8-27B had no 4.1 entry. The other base-model figures also come from Qwen's model card, a paper or Artificial Analysis rather than from RadixArk's runs. Terminal Bench 4 tests are hard by design: Vals AI described 4.0 as long, expert-level terminal jobs graded all or nothing, which helps explain why low scores are common. The smaller gains elsewhere point the same way, though they also pair RadixArk's runs with other publishers' numbers. The doubling on that test compares different benchmark versions and different agent setups, and an independent run on matching conditions would settle how large the gain is.
- RadixArk's DaisyRose-27B is Qwen3.8-27B further trained with reinforcement learning on about 1,200 curated long-horizon agent tasks for 18 days on 144 H200s. Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, with smaller gains on six other reported benchmarks.
- Tasks were redesigned around mistakes the base model actually made, tuned until the solve rate sat in a useful learning range, and reviewed so none could be passed without solving the intended problem. Each attempt got up to 1.5 hours and more than 100 tool calls.
- Training ran fully asynchronously in RadixArk's Miles framework: 128 GPUs ran the agent while 16 trained on its trajectories, with new weights pushed back after every step. RadixArk reports 37% MFU versus 16% for the open-source Megatron trainer, so 57% less GPU time.
- The sandbox service, built on open-source AgentENV, ran about 944,000 sandboxes over nine weeks with up to 1,566 at once. RadixArk says letting each sandbox's memory grow and shrink with use cut infrastructure costs about 80% versus popular cloud sandbox providers.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
Checking sign-in…
Loading comments…


