🔍 Read the full analysis: DaisyRose-27B: Pushing The Frontier Of 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 independent verification and fuller evaluation details are not included in the supplied material.
RadixArk has released DaisyRose-27B, a Qwen3.8-27B model post-trained with reinforcement learning on tasks that require extended planning and tool use, as detailed in the original analysis. The company reports that the model’s Terminal-Bench 4.1 pass@1 score rose from 5.6% to 12.0%; the result is company-reported, and independent verification is not provided in the source material.
RadixArk describes DaisyRose-27B as its first model release and says it was trained using Miles, the company’s reinforcement-learning framework. Its curated training set contained about 1,200 tasks, selected from thousands of candidates developed with Eigent AI. The company says tasks were chosen and revised to target errors made by the original model, including attempts that could otherwise pass without solving the intended problem.
The tasks were designed to require multiple connected steps. According to RadixArk, an attempt could run for up to 1.5 hours, make more than 100 tool calls and generate more than 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.
The release makes BF16, FP8 and NVFP4 checkpoints available under Apache 2.0. RadixArk says DaisyRose-27B retains the base model’s architecture, tokenizer, chat template, image understanding and 256,000-token context window. The company reports higher scores than the original model on all seven listed benchmarks, including Terminal-Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, text-only HLE, IFBench and GPQA Diamond. It says the reported score for each benchmark uses the higher result from BF16 and FP8 runs, which differed by less than one point where both formats were tested.
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 Agent Workflows
The release addresses a practical challenge for AI agents: performing work that involves many steps, tools and extended interaction, rather than producing a short answer in one exchange. If the reported gains hold under independent testing, they could suggest that reinforcement-learning post-training can improve performance on some long-horizon tasks without changing the underlying model architecture.
RadixArk also presents the training system’s resource use as part of the result. It reports 37% model FLOPs utilization for its internal backend, compared with 16% for its open-source trainer on comparable runs, and describes the difference as 57% less GPU time for the same training needs. Including agent-task time, it reports about 16,800 tokens per second across 144 GPUs, versus about 8,300 for the open-source trainer using the same number of GPUs. Those figures may interest organizations weighing the cost of agent training, but their broader relevance depends on the workloads and measurement settings used.
The scores do not establish how the model will perform in everyday deployments. Benchmarks measure particular tasks under particular conditions; real-world reliability, operating cost and performance on work outside those evaluations remain separate questions.
Products Worth Considering

LG UltraGear 34-Inch Curved QHD Gaming Monitor, 160Hz, FreeSync Premium

Bose SoundLink Plus Portable Bluetooth Speaker, Black

Jackery HomePower 3000 Portable Power Station, 3072Wh, 3600W
As an affiliate, we earn on qualifying purchases.
From Qwen3.8 to Post-Training
DaisyRose-27B is presented as a post-trained version of Qwen3.8-27B, not as a newly designed architecture. RadixArk says it used an asynchronous training setup: 128 H200 GPUs handled agent-task inference, while 16 GPUs trained on the resulting trajectories. Updated models were then sent back to the inference GPUs.
The company says the curated tasks were run in 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 at once. RadixArk also reports a 99% sandbox-creation success rate across nine weeks, rising above 99.95% since late October, and says its sandbox infrastructure costs were about 80% lower than those of popular cloud sandbox providers. These operational figures are also claims from the release.
The supplied material describes a training run completed in 18 days using 144 H200 GPUs. It also says an internal-backend run lasted weeks and consumed 63,000 GPU-hours. The source does not explain how those descriptions relate, so they should not be treated as directly reconciled measures of the same run.
““DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.””
— RadixArk
Benchmark Details Still Needed
The supplied release is the basis for the benchmark, throughput, reliability and cost figures in this report. It does not provide an independent evaluation or enough information to assess statistical variation, run-to-run consistency, or the full settings behind each benchmark. RadixArk says its model card includes comparison sources and evaluation settings, but those details were not included in the source material provided here.
It is also unclear how much of the benchmark improvement reflects the reinforcement-learning method, the selected task set, or other choices in evaluation and model use. The release does not reconcile its descriptions of an 18-day training run and a run lasting weeks, nor does the supplied material explain the method behind its claimed sandbox-cost reduction. The company’s figures therefore remain claims pending fuller disclosure and reproduction.
Independent Reproduction Is Key
RadixArk points readers to the Apache 2.0 checkpoints and model card for model properties and inference settings. It 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, using disclosed evaluation settings and checkpoint formats. Further reporting can also clarify the training-run timeline and provide detail on the cost and throughput measurements. Until then, the reported results offer a company account of DaisyRose-27B’s performance rather than independently established evidence of how broadly the gains apply.
Key Questions
What is DaisyRose-27B?
It is a Qwen3.8-27B model that RadixArk says it post-trained with reinforcement learning on long-horizon agent tasks. The company describes it as its first model release.
What benchmark gain did RadixArk report?
RadixArk reports that Terminal-Bench 4.1 pass@1 increased from 5.6% to 12.0% compared with the original model. The supplied material does not include independent confirmation of the result.
What training resources did the company report?
The release says training used 144 H200 GPUs and was completed in 18 days. It also describes a run lasting weeks and consuming 63,000 GPU-hours, without explaining how those accounts fit together.
Can developers access the model?
RadixArk says BF16, FP8 and NVFP4 checkpoints are available under the Apache 2.0 license. It also points to a model card and an SGLang cookbook for model and serving information.
Are the reported results independently verified?
Not in the supplied material. The benchmark and infrastructure figures come from RadixArk’s release. Independent reproduction and fuller evaluation details would help establish how reliably the results generalize.
Primary source: Hugging Face · via ThorstenMeyerAI.com
