🔍 Read the full analysis: Can DaisyRose-27B Improve AI Post-Training Efficiency? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
RadixArk has released DaisyRose-27B, a Qwen3.8-27B model post-trained with reinforcement learning on long-horizon agent tasks. The company reports a Terminal-Bench 4.1 pass@1 score of 12.0%, up from 5.6% for the base model, but the results and efficiency claims have not been independently verified in the supplied material.
RadixArk has released DaisyRose-27B, its first model, and says reinforcement-learning post-training on long-horizon agent tasks raised its Terminal-Bench 4.1 pass@1 score from 5.6% to 12.0%. The checkpoints are available under the Apache 2.0 license, while the benchmark and training-efficiency figures remain company-reported and await independent verification.
The release describes DaisyRose-27B as a Qwen3.8-27B model trained with Miles, RadixArk’s reinforcement-learning framework. The company says it selected about 1,200 tasks from thousands of candidates developed with Eigent AI. Tasks were designed around errors made by the original model and often required several connected actions, including planning and tool use.
According to RadixArk, an individual attempt could last up to 1.5 hours, make more than 100 tool calls and generate more than 256,000 output tokens. Training sessions could span as many as 1 million tokens, with summaries used to carry work across trajectories. The company says it reviewed attempts and removed or revised tasks that could be passed without solving the intended problem.
RadixArk reports higher scores than the base model on all seven benchmarks it lists: Terminal-Bench 4.1 and 2.1, DeepSWE v1.1, SWE Bench Pro v2, text-only HLE, IFBench and GPQA Diamond. The release says each reported score uses the higher result from BF16 and FP8 runs; where both formats were tested, scores differed by less than one point. The company also makes BF16, FP8 and NVFP4 checkpoints available under Apache 2.0.
DaisyRose-27B Trains for Longer Agent Tasks
RadixArk reports benchmark gains after post-training Qwen3.8-27B with reinforcement learning. The results are company-reported and need independent verification.
What changed
The release describes post-training an existing model foundation, not a new architecture.
Built on Qwen3.8-27B
RadixArk says it retained the architecture, tokenizer, chat template, image understanding and 256,000-token context window.
Long-horizon reinforcement learning
Tasks required connected steps such as planning and tool use, with Miles handling reinforcement learning.
How tasks stretched the model
Attempts could span extended interactions and carry work across multiple trajectories.
About 1,200 tasks were selected from thousands of candidates and refined around errors made by the original model.
An attempt could last up to 1.5 hours, make more than 100 tool calls and produce more than 256,000 output tokens.
The asynchronous setup used 128 GPUs for task inference and 16 GPUs for training.
Reported efficiency
RadixArk compares its internal backend with its open-source trainer on comparable runs.
What readers can check
The release offers checkpoints and points to the model card for properties and inference settings.
- BF16, FP8 and NVFP4 checkpoints
- Apache 2.0 license
- SGLang and DSpark serving recommendations
- Independent benchmark reproduction
- Repeated-run and evaluation details
- Reconciliation of 18 days with weeks-long training and 63,000 GPU-hours
What the Agent Benchmarks Measure
The release addresses a practical challenge in AI development: training agents to complete tasks that demand many steps, tool calls and extended interaction, rather than producing one short answer. A higher score on Terminal-Bench 4.1 could indicate better performance on the kinds of terminal-based tasks the benchmark tests, if the result holds under independent evaluation.
RadixArk also presents training cost as part of the announcement. It reports 37% model FLOPs utilization for its internal backend, compared with 16% for an open-source trainer on comparable runs, and says this represented 57% less GPU time for the same training needs. Including agent-task time, it reports about 16,800 tokens per second across 144 GPUs, compared with about 8,300 for the other trainer on the same number of GPUs. Those figures could interest teams considering extended agent training, but their relevance depends on the comparison setup and whether other groups can reproduce them.
The reported gains across several benchmarks suggest the company’s claim is not confined to Terminal-Bench 4.1. They do not, by themselves, establish how DaisyRose-27B will perform in everyday use or show that reinforcement learning alone caused the differences.
Products Worth Considering

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Logitech MX Brio Ultra HD 4K Webcam

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten
As an affiliate, we earn on qualifying purchases.
From Qwen3.8 to Agent Training
DaisyRose-27B is a post-trained model, not a new architecture, according to RadixArk. The release says it retains Qwen3.8-27B’s architecture, tokenizer, chat template, image understanding and 256,000-token context window. The model card is identified as the source for its properties and inference settings.
RadixArk says training ran asynchronously: 128 H200 GPUs handled agent-task inference while 16 GPUs trained on the resulting trajectories, with updated models sent back to the inference machines. The company describes the run as lasting 18 days in its summary, but elsewhere says an internal-backend run lasted weeks and consumed 63,000 GPU-hours. The supplied material does not explain how those descriptions fit together.
The company also reports that its AgentENV-based sandbox service completed about 944,000 sandbox runs between mid-September and mid-November, with as many as 1,566 running at once. It says 99% of sandbox creations succeeded over nine weeks, rising above 99.95% since late October, and claims infrastructure costs were about 80% lower than with popular cloud sandbox providers. These are vendor-reported figures; the supplied release does not give enough detail to assess the comparisons independently.
““DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.””
— RadixArk
Benchmark and Cost Claims Await Checks
The benchmark, speed, reliability and cost figures in this article come from RadixArk’s release. The supplied information does not include an independent evaluation or enough detail to establish statistical variation, run-to-run consistency, or the full settings used for every benchmark. The company says its model card contains evaluation sources and settings, but those details were not included in the source material.
It is also unclear how much of the reported benchmark improvement is attributable to the training method, task selection or other evaluation and model-use choices. The mismatch between the summary’s 18-day training description and the account of a run lasting weeks and using 63,000 GPU-hours is unresolved. The source’s explanation of how sandbox infrastructure costs were reduced is incomplete as well.
Independent Reproduction Is the Test
RadixArk points readers to the Apache 2.0 checkpoints and model card for model properties and inference settings. It recommends serving DaisyRose-27B with SGLang and DSpark speculative decoding and says its SGLang cookbook includes launch configurations for different GPUs and checkpoint formats.
The next meaningful test is whether independent evaluators can reproduce the benchmark comparisons using clearly documented settings and checkpoint formats. Further detail from RadixArk on the training-duration discrepancy and the basis for its cost comparisons would also help readers judge the efficiency claims. No independent reproduction or additional results are included in the supplied material.
Key Questions
What is DaisyRose-27B?
It is RadixArk’s Qwen3.8-27B model post-trained with reinforcement learning on long-horizon agent tasks. The company says the checkpoints are available under the Apache 2.0 license.
What result does RadixArk report on Terminal-Bench 4.1?
RadixArk reports a pass@1 score of 12.0%, compared with 5.6% for the original model. The result is company-reported and has not been independently verified in the supplied material.
How long did training take?
The release summary says 18 days using 144 H200 GPUs. Elsewhere it describes a run lasting weeks and using 63,000 GPU-hours; the available information does not reconcile those descriptions.
Are the efficiency claims independently confirmed?
No independent confirmation is provided in the source material. RadixArk reports better GPU utilization and throughput than an open-source trainer, but the comparison settings and reproducibility remain important to evaluating those claims.
Primary source: Hugging Face · via ThorstenMeyerAI.com
