- 6 min read
- 8 key claims
- 1 source cited
- For Developers, Businesses
Overview
RadixArk has published DaisyRose-27B, its first model, built on Qwen3.8-27B and post-trained with reinforcement learning through the Miles framework [1]. The operational recipe is the main contribution: tasks were redesigned around failures observed in the base model; 128 H200 GPUs ran agent inference while 16 trained the model asynchronously; sessions reaching one million tokens were folded into multiple trajectories through harness summarization; and each attempt executed in a fresh, disposable sandbox [1].
The resulting checkpoints are reported under the Apache 2.0 license and preserve Qwen3.8-27B's architecture, tokenizer, chat template, image-understanding capability and 256K-token context window [1]. The model reasons before answering by default, exposing reasoning in reasoning_content and the final response in content [1].
Why it matters
Agent post-training is constrained by several systems at once: difficult task design, long-running inference sessions, expensive accelerator time and disposable execution environments. DaisyRose-27B describes how RadixArk connected those pieces in one run [1].
The developer-reported results include a Terminal-Bench 4.1 pass@1 increase from 5.6% for the base model to 12.0% for DaisyRose-27B, using the higher of the BF16 and FP8 runs for each benchmark [1]. That is a measured improvement in the developer's account, but it is not an independently reproduced result, and the base-model naming is not independently contextualized in the available material [1]. The release also retains the stated base model's interface and publishes checkpoints under an open license, so teams can inspect the artifact and adapt it to their own workflows [1].
Opportunities
- Inspect and adapt the checkpoint. Teams using the stated Qwen3.8-27B interface can test the checkpoints directly rather than rebuilding the serving and chat-template integration [1].
- Study asynchronous reinforcement learning. The split between 128 inference GPUs and 16 trainer GPUs is a concrete reference for pipelines where long agent sessions would otherwise idle a trainer [1].
- Borrow failure-driven task design. Running the current model against candidate tasks, redesigning them around observed mistakes and removing tasks passable without solving the intended problem is a methodology teams could apply to proprietary workflows [1].
- Plan for long sessions deliberately. The combination of a 256K context window and harness summarization across trajectories offers one approach to agent training that exceeds the native context limit [1].
- Use the open components as references. Miles, the SGLang launch configurations and the AgentENV-based sandbox service are described as open-source foundations for further work [1].
Risks
- The reported gains are unverified. All visible benchmark, utilization, throughput, stability and cost figures come from RadixArk's own account, with no third-party test report [1].
- Benchmark coverage is narrow. Terminal-Bench 4.1 is the only visible score; scores for the other reasoning, instruction-following and coding benchmarks are not included, and no independent assessment covers general capability or regressions [1].
- A 12.0% pass@1 rate leaves substantial task failure. The improvement is relative to a 5.6% baseline rather than evidence of reliable long-horizon completion [1].
- Infrastructure reproduction is uncertain. The comparison settings, trainer version, accounting boundaries and sandbox pricing assumptions are not fully specified [1].
- The compute footprint is substantial. The developer reports 63,000 GPU-hours on 144 H200s for the release run, before any replication or experimentation by an adopting team [1].
Context
DaisyRose-27B targets long-horizon agent work rather than a narrow chat benchmark. RadixArk and Eigent AI generated thousands of candidate tasks across machine learning, software engineering and science; about 1,200 entered the final set [1].
The method starts with observed failures. RadixArk ran Qwen3.8-27B on every candidate and redesigned tasks around mistakes it wanted the model to stop making, then adjusted difficulty until the solve rate fell in what the developer calls a useful learning range [1]. The published account does not include a specific before-and-after task, so the concrete shape of a redesigned task is unknown [1]. Each task requires several interdependent steps, with up to 1.5 hours in a sandbox, more than 100 tool calls and more than 256K output tokens per attempt [1].
Because training sessions can reach one million tokens while the model retains a 256K context window, the pipeline uses harness summarization across trajectories rather than relying on one native-context rollout [1]. The developer's model card lists a selected serving stack of SGLang with DSpark speculative decoding, including launch configurations, but that stack is not presented as a requirement of the checkpoint [1].
Known, claimed, inferred
The release is best read as an operational recipe rather than a demonstrated step change in model quality per unit of compute. Its distinctive combination is failure-driven task design, asynchronous separation of inference from training, long-session summarization and disposable sandbox execution [1].
RadixArk reports that its internal backend reached 37% model FLOPs utilization, compared with 16% for an open-source Megatron-based trainer on comparable runs, and translates that into 57% less GPU time for the same training [1]. Counting agent task time, the release run reportedly processed about 16,800 tokens per second across 144 GPUs against about 8,300 for the open-source trainer on the same GPUs [1]. These are developer-reported GPU-time and throughput improvements. The comparison conditions, trainer version and accounting boundaries are not fully specified, and the record does not establish lower end-to-end training cost or better model quality per unit of compute [1].
The same account reports larger gains for other model families: 6.2× on Qwen3.6-35B-A3B and more than 10× on the roughly 300B-parameter GLM-5.3-Flash [1]. These are developer-reported observations about mixture-of-experts models, not DaisyRose-27B, and they do not establish behavior for this release [1].
Limitations & uncertainties
All visible results come from one official account. The Terminal-Bench 4.1 figure selects the higher of the BF16 and FP8 runs for each benchmark, where both were evaluated they differed by less than one point; this is a developer-reported reporting choice rather than an independently reproduced result [1]. Scores for other reasoning, instruction-following and coding benchmarks, along with the referenced comparison sources and evaluation settings, are not visible [1].
The efficiency comparisons do not fully specify the open-source trainer version, the exact comparable-run setup, or whether token-count and elapsed-time boundaries were identical. The meaning of model FLOPs utilization for this workload is undefined, and the effect of harness summarization on token accounting and quality is not reported [1]. The sandbox-cost comparison does not specify the hardware, configuration or pricing assumptions used [1].
The sandbox figures describe RadixArk's own environment for this project over nine weeks, not a generally available service: about 944,000 sandboxes created, up to 1,566 concurrent, 99% creation success across the period and above 99.95% since late October [1]. The published text is truncated, leaving qualifications after the sandbox-memory discussion unknown [1].
Dataset composition, annotation protocol, reward design, task-selection criteria beyond the mistake-driven process and safeguards against overfitting are not described. No independent results cover general capability, task reliability, inference behavior or possible regressions [1]. The base-model name Qwen3.8-27B is not independently contextualized, so its provenance and the accuracy of its naming are unestablished [1].
The checkpoints are reported as Apache 2.0-licensed, but that statement does not establish the status of the base model, the training data, generated task content or third-party sandbox code [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementDaisyRose-27B was built on Qwen3.8-27B and post-trained with reinforcement learning using Miles; its checkpoints are reported under Apache 2.0 and retain the stated base model's architecture, tokenizer, chat template, image understanding and 256K context window.[1]
- Primary-source or vendor statementThe final training set contained about 1,200 tasks selected from thousands of candidates and redesigned around mistakes observed in the base model, with difficulty adjusted to a developer-defined learning range.[1]
- Primary-source or vendor statementRadixArk reports Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0%, using the higher of the BF16 and FP8 runs for each benchmark; this is a developer-reported selection rule, not an independently reproduced result.[1]
- Primary-source or vendor statementTraining ran asynchronously, with 128 H200 GPUs running agent inference and 16 running the trainer, while sessions reaching one million tokens were folded into multiple trajectories through harness summarization.[1]
- Primary-source or vendor statementRadixArk reports 37% model FLOPs utilization on its internal backend versus 16% for an open-source Megatron-based trainer on comparable runs, and about 16,800 tokens per second versus about 8,300; the comparison conditions, trainer version and accounting boundaries are not fully specified.[1]
- Primary-source or vendor statementThe release run consumed a reported 63,000 GPU-hours on 144 H200s over 18 days, with stable speed and efficiency according to the developer.[1]
- Primary-source or vendor statementRadixArk's own sandbox environment for this project ran about 944,000 sandboxes over nine weeks, with up to 1,566 concurrent and 99% creation success, rising above 99.95% since late October.[1]
- Primary-source or vendor statementThe reported MoE improvements of 6.2× on Qwen3.6-35B-A3B and more than 10× on the roughly 300B-parameter GLM-5.3-Flash concern other model families and do not establish behavior for DaisyRose-27B.[1]
Sources
- 1DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency●huggingface.coretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.