What DaisyRose-27B Reveals About Efficient Agent Post-Training
RadixArk’s DaisyRose-27B release pairs a compact, curated task set with asynchronous reinforcement learning and long sandbox trajectories. The useful lesson is less about one leaderboard gain than about how teams can evaluate the full post-training system.

Efficient agent post-training depends on curated failure-focused tasks, reliable sandboxes, and end-to-end pipeline measurement—not trainer speed or benchmark scores alone.
RadixArk has released DaisyRose-27B, a Qwen3.8-27B-based model post-trained with reinforcement learning on long-horizon agent tasks. The release is notable not simply because a 27B-parameter checkpoint improved on its base model, but because it exposes several interacting parts of an agent-training pipeline: task construction, asynchronous learning, sandbox reliability, and evaluation.
The team reports training on roughly 1,200 curated tasks over 18 days using 144 H200 GPUs. Its headline result on Terminal-Bench 4.1 rises from 5.6% pass@1 for the base model to 12.0%, alongside smaller reported gains on six other reasoning, coding, and instruction-following evaluations. BF16, FP8, and NVFP4 checkpoints are available under Apache 2.0.
Why a small curated set can still matter
A count of 1,200 tasks sounds modest next to the enormous corpora used for pretraining. That comparison is misleading. Post-training examples can be reused through multiple rollouts, and each long agent attempt produces a sequence of decisions, tool interactions, failures, and rewards. The effective learning signal depends on trajectory quality and diversity, not only the number of task prompts.
The release describes a deliberate selection loop: begin with problems the base model gets wrong, tune their difficulty into a learnable range, require multiple dependent steps, and remove tasks that can earn reward without completing the intended work. This is a practical defense against two common problems. Tasks that are too easy provide little corrective signal; tasks that are effectively impossible mostly generate uninformative failure. Meanwhile, weak graders can teach the model to exploit the evaluation rather than solve the problem.
For teams building a similar curriculum, task labels should therefore capture more than domain. Useful metadata includes the base model's failure mode, expected tool sequence, dependency depth, grader confidence, and observed solve rate. Reviewing these attributes helps distinguish genuine skill acquisition from memorizing recurring task shapes.
The system, not just the trainer, sets throughput
Long-horizon reinforcement learning couples two very different workloads. Agents generate trajectories through inference and external tools, while training workers update model weights from completed experience. If either side waits frequently for the other, expensive accelerators can sit idle.
DaisyRose used the open-source Miles framework in a fully asynchronous arrangement, with 128 H200s assigned to agent inference and 16 to training. RadixArk reports 63,000 GPU-hours for the release run and about 16,800 aggregate training tokens per second. It also reports 37% model FLOPs utilization on its internal backend, compared with 16% for its open-source trainer on comparable runs. Those figures describe the team's particular hardware and workload; they should not be treated as universal efficiency expectations.
The broader design lesson is to measure the pipeline at its boundaries. Trainer utilization alone does not reveal time lost to sandbox startup, task execution, grading, failed rollouts, weight distribution, or queues between stages. A useful operational dashboard would separate trajectory generation time, accepted-token throughput, update time, infrastructure failure rate, and end-to-end cost per usable trajectory. That breakdown makes bottlenecks actionable.
Long context is not the same as long-horizon competence
The model retains its base architecture and 256K-token context window, while some training sessions reportedly reached one million tokens through harness summarization across trajectories. This distinction matters. A model's native context limit describes how much it can attend to in one invocation. A harness can support a much longer job by compressing prior work and feeding a summary into later calls, but information may be lost or distorted at each compression step.
Evaluation should consequently test the complete loop. Can the agent preserve requirements after several summaries? Does it retain the location and meaning of earlier artifacts? Can it recognize when a summary omitted a critical constraint? Success on short coding questions will not answer those questions. Representative tests should include interrupted workflows, misleading tool output, partial failures, and tasks whose acceptance criteria are established early but checked much later.
How to evaluate the released checkpoints
The reported benchmark improvements are encouraging, yet the 12.0% Terminal-Bench 4.1 result also means most attempts under that setup did not pass. Teams considering the model should start with a small evaluation set drawn from their actual workflows and preserve the serving and decoding configuration used for each run. The model card and serving guide should be treated as the configuration baseline, not an optional detail.
Compare checkpoints on task completion, but also record tool-call count, latency, token use, invalid actions, recovery from errors, and variance across repeated attempts. Because the published table uses the better of BF16 and FP8 results where both were evaluated, local comparisons should avoid silently selecting the best format after observing outcomes. Choose the deployment format in advance or report every tested format separately.
Finally, inspect failures rather than relying on an aggregate score. A model that improves average completion while becoming less reliable at permission boundaries, destructive commands, or requirement retention may be a poor fit for production automation. Sandboxing and least-privilege tools remain necessary regardless of benchmark performance.
A reproducible release with clear questions left open
DaisyRose-27B makes several useful artifacts available: multiple checkpoint formats, an open license, a serving recipe, and the Miles training framework. That gives practitioners concrete material to inspect and test. At the same time, the internal backend efficiency figures and the project's infrastructure scale will not automatically transfer to smaller clusters or different agent environments.
The durable contribution is a systems-oriented view of post-training. Better agent behavior can emerge from carefully targeted tasks, reliable graders, efficient scheduling, and robust execution environments working together. The next step for adopters is not to assume the same gains, but to reproduce the evaluation on their own workload and identify which part of that system provides the value.
Source: DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency ↗. How we write


