RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned for agent tasks

RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned for agent tasks

RadixArk has released DaisyRose-27B, its first model. It is built on Qwen3.8-27B and trained with reinforcement learning using Miles, RadixArk's open-source RL framework for large-scale post-training. The training used long-horizon agent tasks with sessions of up to 1M tokens: just 1,200 curated tasks, 18 days, 144 H200 GPUs. RadixArk says DaisyRose-27B lifts Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%, a gain of 6.4 percentage points on what the post calls a notoriously hard long-horizon benchmark. It also says the model scores higher than the original Qwen3.8-27B on every benchmark in its chart (Terminal Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, HLE, IFBench, GPQA Diamond), with the largest gain on Terminal Bench 4.1. The chart's other scores are not reproduced in the post's text. BF16, FP8 and NVFP4 checkpoints are available on Hugging Face.

RadixArk credits three ingredients. The first is training data. The tasks were built with Eigent AI, starting from its large pool of tasks in machine learning, software engineering, science and other fields. Thousands of candidates were built together, and about 1,200 reached the final set. Each task had to meet three criteria. It targets a mistake the original Qwen3.8-27B makes: the team ran the base model on every task, redesigned it to bring out that mistake, and tuned it until the solve rate was neither too easy nor too hard to learn from. Every task is a long job with several chained, interdependent problems, and each attempt gets up to 1.5 hours in the sandbox, with more than 100 tool-calls and more than 256k output tokens. And there is no room for reward hacking: every task and the model's attempts at it were reviewed, and any task that could be passed without really solving it, or that had a design flaw, was fixed or removed.

The second ingredient is the training backend. Miles ran in fully asynchronous mode, so the agent keeps working while the model learns and neither waits for the other. 128 H200s ran inference on agent tasks and 16 ran the trainer, with the updated model sent to the 128 GPUs after every step. Learning ran on RadixArk's internal backend, built for agent training: each step learns from a handful of very long sessions, some as long as 1M tokens, folded into multiple trajectories through harness summarization. On this release, 37% of each GPU's peak compute went into useful work (model FLOPs utilization, MFU), against 16% for the open-source Megatron-based trainer on comparable runs. RadixArk puts that at 57% less GPU time for the same training. It reports larger gains on mixture-of-experts models: 6.2x on Qwen3.6-35B-A3B and more than 10x on the roughly 300B-parameter GLM-5.3-Flash. Counting the time the agent spends on its tasks, the release run trained at about 16,800 tokens per second across its 144 GPUs, against about 8,300 for the open-source trainer on the same GPUs. The run lasted weeks, 63,000 GPU-hours in total, and speed and efficiency held steady from first step to last.

The third ingredient is the sandbox service. Each agent attempt runs in a sandbox, a fresh copy of the task's environment where the agent runs commands, edits files and is then graded. RadixArk ran them on its own service, built on the open-source AgentENV project, to which it also contributes fixes. Between mid-September and mid-November the service ran about 944,000 sandboxes for this project, with up to 1,566 running at once. It grew from one machine to as many as 12, and ten upgrades shipped, nearly all while training was running. Across nine weeks, 99% of sandbox creations succeeded, and since late October more than 99.95%. One fix was keeping a machine from freezing when it runs out of memory. RadixArk says its cloud cost for the project's sandboxes was about 80% lower than running the same sandboxes on popular cloud sandbox providers. One reason is memory that grows and shrinks with actual use: a sandbox usually uses only a small part of what it asks for, so each machine can run many more sandboxes.

RadixArk says it is bringing this approach to teams building specialized models, aiming to improve the capabilities they need while preserving the model's broader strengths. The post thanks Eigent AI, the Qwen team, the Miles and SGLang communities, and open-source projects including slime, AgentENV and Harbor.

Key facts

  • DaisyRose-27B is RadixArk's first model, built on Qwen3.8-27B and trained with RL using the open-source Miles framework; BF16, FP8 and NVFP4 checkpoints are on Hugging Face.
  • RadixArk reports Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0% (8 runs per task, internal mini-swe-agent harness with summarization), and higher scores than the base model on every benchmark in its chart.
  • Training used about 1,200 curated tasks built with Eigent AI, sessions of up to 1M tokens, 18 days on 144 H200 GPUs; the release run totalled 63,000 GPU-hours.
  • RadixArk's internal backend hit 37% MFU versus 16% for the open-source Megatron-based trainer, which it states as 57% less GPU time, and about 16,800 versus 8,300 tokens per second end to end.
  • Its sandbox service ran about 944,000 sandboxes, with up to 1,566 at once and more than 99.95% creation success since late October; RadixArk says sandbox cost was about 80% below popular cloud providers.

Why it matters

The release is a worked example of agent-focused reinforcement learning on a modest budget: 1,200 tasks, 144 H200s and 18 days moved a 27B model on a hard long-horizon terminal benchmark from 5.6% to 12.0%. RadixArk's argument is that carefully designed tasks and efficient infrastructure, not sheer scale, produced the gain. The post also puts figures on the infrastructure side: training efficiency, sandbox reliability and sandbox cost. Those are the parts of agent RL that are often hard to see from the outside.

Who it affects

Teams that want an open agent-oriented model built on Qwen3.8-27B can download and build on the checkpoints. Groups doing RL post-training can look at Miles, the open-source framework RadixArk used. Companies running large numbers of agent sandboxes will care about the sandbox cost and reliability claims. RadixArk also says it is offering this approach to teams building specialized models and invites them to contact it.

How to use it

DaisyRose-27B can be downloaded from Hugging Face in BF16, FP8 and NVFP4 checkpoints. Teams that want to train their own agents can use Miles, RadixArk's open-source reinforcement learning framework. The learning in this release ran on RadixArk's internal training backend rather than the open-source trainer, and the sandboxes ran on RadixArk's own service built on AgentENV. For help with post-training a specialized model, the post gives contact@radixark.ai. No license for the weights is stated in the post.

How solid is it

This is a first-party announcement, and every number is RadixArk's own. No independent evaluation is mentioned. The headline score is an average over 8 runs per task using RadixArk's internal mini-swe-agent harness with summarization turned on; the other benchmarks use stated run counts and harnesses, and HLE is graded by GPT-6 Luna with no tools. The post gives scores only for Terminal-Bench 4.1, so the size of the gains elsewhere is unknown. The post does not say whether 12.0% is competitive with other models, only that it improves on the base model's 5.6%. Its efficiency comparisons are against the open-source Megatron-based trainer on comparable runs, and the 80% sandbox saving names no specific cloud providers.

Risks and caveats

An absolute score of 12.0% is still low, and the post frames it only as a gain over the 5.6% base. The training figures are not fully reconciled: the intro says 18 days, a later section says the run trained for weeks. The 57% savings and the 2x speedup depend on RadixArk's internal backend, which is not the open-source trainer, so Miles users should not assume they will get the same numbers. The post gives no total training cost in money. The tasks were designed around mistakes the base model makes, so the gains may track those targeted tasks, and the post says only that the model scores higher on every benchmark shown.

“Our magic lies in the following three key ingredients: high-quality training data, efficient training backend, and robust sandboxes.”

— RadixArk, DaisyRose-27B announcement