RadixArk Releases DaisyRose 27B, Its First Open Weights Model, Trained With RL for Long Agent Tasks
How much can you improve a small model in 18 days on 144 GPUs? Meet DaisyRose-27B, RadixArk's first open-weight model! It is trained with RL on Qwen3.8-27B, with 1,200 curated long-horizon agent tasks for 18 days on 144 H200s. Powered by Miles and our own training backend, we stretched agent sessions up to 1M tokens and raised the Terminal-Bench 4.1 score from 5.6% to 12.0%. See how we pull it off ⬇️
Context
It shows a small team can roughly double an open model's agent benchmark score with 18 days of reinforcement learning on 144 GPUs, a recipe other builders can study and copy.
RadixArk posted that DaisyRose 27B is its first open weights model, trained with reinforcement learning on top of Qwen3.8 27B. The company said it used 1,200 curated long horizon agent tasks, trained for 18 days on 144 H200 GPUs, and stretched agent sessions up to 1 million tokens. It said the training raised the model's Terminal Bench 4.1 score from 5.6% to 12.0%.
Deep dive
RadixArk posted that DaisyRose 27B is its first open weight model. The company framed the release around a question: how much can you improve a small model in 18 days on 144 GPUs?
According to RadixArk, DaisyRose 27B is trained with reinforcement learning on Qwen3.8 27B, an open model from Alibaba's Qwen team. The company said it used 1,200 curated long horizon agent tasks and trained for 18 days on 144 Nvidia H200 GPUs. It said the work was powered by Miles and its own training backend, and that it stretched agent sessions up to 1 million tokens.
The main result RadixArk reported is on Terminal Bench 4.1, a benchmark of agent tasks in a computer terminal, where it said the score rose from 5.6% for the base model to 12.0% after training. That is more than double the starting score, though still a low absolute number, and the figures come from RadixArk itself rather than an independent evaluation.
The post points readers to further details on how the team did it. The release fits a wider trend of smaller labs taking open Qwen models and adding focused reinforcement learning for agent work, then publishing the weights so others can build on them.
What to expect next
Watch for independent tests of DaisyRose 27B on Terminal Bench and other agent benchmarks to see whether the gains hold. Developers will likely compare it with the base Qwen3.8 27B and other agent tuned open models of similar size. RadixArk may share more on its training data and backend. A bigger follow up model, or the same recipe applied to other base models, would show whether the approach scales.
