🔍 Read the full analysis: A Closer Look At DaisyRose-27B’s AI Post-Training Methods on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
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 increase from 5.6% to 12.0%, but the supplied release does not include independent verification or enough evaluation detail to assess the result fully.
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%, as described in the original analysis. The checkpoints are available under the Apache 2.0 license, but the results are company-reported and independent verification and fuller evaluation details are not included in the supplied material.
RadixArk says DaisyRose-27B is based on Qwen3.8-27B and was trained with Miles, the company’s reinforcement-learning framework. The release describes a curated set of about 1,200 tasks, selected from thousands of candidates developed with Eigent AI. According to RadixArk, the tasks were designed around errors made by the original model and required connected steps such as planning and tool use.
The company says an individual attempt could run for up to 1.5 hours, make more than 100 tool calls and generate over 256,000 output tokens. Longer training sessions could span as many as one million tokens, with harness summarization used to carry work across multiple trajectories. RadixArk says it reviewed attempts and removed or revised tasks that could be passed without solving the intended problem.
RadixArk reports higher scores than the original model across all seven benchmarks listed in its release: Terminal-Bench 4.1 and 2.1, DeepSWE v1.1, SWE Bench Pro v2, text-only HLE, IFBench and GPQA Diamond. The company says it reports the higher result from BF16 and FP8 runs for each benchmark; where both formats were tested, their scores differed by less than one point. It offers BF16, FP8 and NVFP4 checkpoints 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
Testing Longer-Running AI Agents
The release addresses a practical challenge in building AI agents: performance on work that takes many steps, uses tools and continues over extended interactions. A model’s ability to handle those tasks may affect whether it can complete multi-stage workflows, rather than only respond to a short prompt. RadixArk’s reported Terminal-Bench gain is its largest among the comparisons listed, while the company also reports smaller gains on six other benchmarks.
Those figures are relevant to developers considering models for agent tasks, but they do not establish how DaisyRose-27B will perform in everyday deployments. Benchmark results depend on task selection, evaluation settings and model use, and the supplied material does not include independent testing. The release therefore presents a result to investigate, not proof that the training method will produce similar gains in other systems or settings.
RadixArk also frames the training infrastructure as part of the announcement. It reports 37% model FLOPs utilization on its internal backend, compared with 16% for its open-source trainer on comparable runs, and says this meant 57% less GPU time for the same training needs. Including agent task time, it reports about 16,800 tokens per second across 144 GPUs, against about 8,300 for the open-source trainer on the same number of GPUs. These are company comparisons; the supplied material does not provide independent measurements or enough detail to determine how broadly they apply.
Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Bose QuietComfort Wireless Noise Cancelling Headphones, Sandstone

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
From Qwen Base to Post-Training
DaisyRose-27B is described as post-training an existing model foundation, rather than introducing a new architecture. RadixArk says it retains the Qwen3.8-27B architecture and tokenizer, along with the chat template, image understanding and a 256,000-token context window. The model card is identified by the company as the source for model properties and inference settings.
RadixArk says the training was fully asynchronous: 128 H200 GPUs handled agent-task inference while 16 GPUs trained on the resulting trajectories, with updated models sent back to the inference GPUs after each step. The release also describes a run lasting 18 days on 144 H200 GPUs. Elsewhere, it says a run on the internal backend lasted weeks and used 63,000 GPU-hours. The supplied material does not explain how these duration descriptions relate.
For task execution, the company says it used fresh sandboxes managed by a service based on the open-source AgentENV project. It reports about 944,000 sandbox runs between mid-September and mid-November, with as many as 1,566 running simultaneously. RadixArk says 99% of sandbox creations succeeded across nine weeks, rising above 99.95% since late October. It also claims the infrastructure cost about 80% less than popular cloud sandbox providers; supporting cost calculations are not included in the supplied material.
““DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.””
— RadixArk
Benchmark Details Still Needed
The benchmark, speed, reliability and cost figures in this article come from RadixArk’s release. The supplied material does not include independent results, statistical variation or run-to-run consistency data. It says the model card contains comparison sources and evaluation settings, but those details were not provided here. The extent to which the benchmark gains stem from the training approach, task selection or other evaluation choices is also unclear.
Two parts of the training account need clarification: the relationship between the stated 18-day run and the description of training lasting weeks, and the basis for the 63,000 GPU-hour figure. The source also makes a cost-reduction claim about its sandbox service without explaining the calculation in the material provided. Those gaps limit how readily readers can compare the reported results with other models or training runs.
Independent Reproduction and Model Access
RadixArk says researchers and developers can examine the Apache 2.0 checkpoints and consult the model card for model properties and inference settings. The company recommends serving the model with SGLang and DSpark speculative decoding and says its SGLang cookbook includes launch configurations for different GPUs and checkpoint formats.
The next useful test is independent reproduction of the benchmark comparisons, with evaluation settings and checkpoint formats made explicit. Further information on the training timeline and the company’s efficiency and sandbox-cost calculations would also help readers judge the claims. The supplied source does not announce a date for such testing or additional disclosures.
Key Questions
What is DaisyRose-27B?
DaisyRose-27B is a model RadixArk says it built on Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks. The company says the released checkpoints are available under Apache 2.0.
What benchmark result did RadixArk report?
RadixArk reports that Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0% compared with the original model. This is a company-reported result; independent verification is not included in the supplied material.
How was the model trained?
The release says the model was trained with RadixArk’s Miles reinforcement-learning framework on a curated set of about 1,200 agent tasks. The tasks were intended to require multiple connected steps, including planning and tool use.
Are the reported gains independently verified?
No independent evaluation is included in the source material. RadixArk says its model card contains evaluation settings and comparison sources, but those details were not supplied here.
What remains unclear about the training run?
The release says training took 18 days, but also describes a run lasting weeks and using 63,000 GPU-hours. It does not reconcile those descriptions, and the supplied material does not fully explain the evaluation or cost calculations.
Primary source: Hugging Face · via ThorstenMeyerAI.com
