🔍 Read the full analysis: Inside DaisyRose-27B: Exploring Efficient Post-Training 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 Terminal-Bench 4.1 pass@1 improved from 5.6% to 12.0%, but independent verification and full evaluation details are not available in the supplied release 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 company has made checkpoints available under the Apache 2.0 license; the reported benchmark results have not been independently verified in the source material provided.
DaisyRose-27B is based on Qwen3.8-27B and was post-trained using 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 mistakes made by the base model and involved multiple connected steps, including planning and tool use.
According to RadixArk, an individual attempt could run for up to 1.5 hours, make more than 100 tool calls and generate more than 256,000 output tokens. Some longer training sessions could span as many as one million tokens, with harness summarization 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 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. It says the comparisons use the higher result from BF16 and FP8 runs for each benchmark and that, where both formats were tested, scores differed by less than one point. The release provides BF16, FP8 and NVFP4 checkpoints.
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 question for AI developers: whether post-training can improve performance on tasks requiring many steps, extended interaction and repeated tool use, rather than a single short response. RadixArk’s reported Terminal-Bench 4.1 result is its largest listed gain, while the company says the model also scored higher on six other benchmarks. Those results, if reproduced under comparable conditions, could inform how teams assess models for software and other agent-style work.
RadixArk also presents the training process as an efficiency result. It says its internal backend reached 37% model FLOPs utilization, compared with 16% for its open-source trainer on comparable runs, and characterizes this as 57% less GPU time for the same training needs. Including agent-task time, the company reports about 16,800 tokens per second across 144 GPUs, compared with about 8,300 for the open-source trainer on the same GPU count. These are company-reported comparisons, not independently confirmed measurements, and their wider relevance depends on the exact workloads and settings.
The Apache 2.0 release allows researchers and developers to inspect and test the checkpoints under that license. Access could make it easier to compare the reported results with other models, but open weights alone do not establish that the training procedure or benchmark outcomes will be reproducible.
Products Worth Considering

DJI Mini 4K Drone Bundle with Strobe Light, Landing Gear, and Propeller Holder Strap

Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten
As an affiliate, we earn on qualifying purchases.
Post-Training an Existing Model
DaisyRose-27B is described as a post-trained version of an existing model, not a new architecture. RadixArk says it retains Qwen3.8-27B’s architecture, tokenizer and chat template, as well as image understanding and a 256,000-token context window. The company points to its model card for model properties and inference settings.
The release 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 group. It also describes a sandbox service based on the open-source AgentENV project. RadixArk reports about 944,000 sandbox runs from mid-September to mid-November, with as many as 1,566 running at once. It says sandbox creation succeeded in 99% of cases across nine weeks, rising above 99.95% since late October, and claims infrastructure costs were about 80% lower than those of popular cloud sandbox providers.
The company’s material describes the training as completed in 18 days, but also says a run on its internal backend lasted weeks and used 63,000 GPU-hours. The supplied information does not explain how those durations relate. The timing of the model’s public release is also not specified in the source material.
““DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.””
— RadixArk
Benchmark and Runtime Questions
The supplied release is the source for the benchmark, speed, reliability and cost figures. It does not provide an independent evaluation or enough information to assess run-to-run variation, statistical uncertainty or every benchmark’s full settings. RadixArk says its model card contains comparison sources and evaluation settings, but those details are not included in the material provided for this report.
It is also unclear how much of the reported benchmark gain came from reinforcement learning itself, task selection, evaluation choices or other aspects of model use. The discrepancy between the 18-day training summary and the description of weeks and 63,000 GPU-hours remains unexplained. The supplied account of the sandbox cost comparison is incomplete, so its calculation and scope cannot be assessed here.
Reproduction and Model Testing
The next useful test is an independent reproduction of the benchmark comparisons, with the evaluation settings, checkpoint formats and inference configuration reported in enough detail for comparison. Researchers can access the Apache 2.0 checkpoints and consult RadixArk’s model card for properties and inference settings, according to the company.
RadixArk 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 supplied material does not name a date for an independent evaluation or a further release milestone. Until such checks are available, the reported gains and efficiency figures should be treated as company claims pending external confirmation.
Key Questions
What is DaisyRose-27B?
It is a model based on Qwen3.8-27B that RadixArk says it post-trained with reinforcement learning on long-horizon agent tasks.
What benchmark improvement did RadixArk report?
The company reports that Terminal-Bench 4.1 pass@1 increased from 5.6% to 12.0%. It also reports higher results than the base model on six other listed benchmarks.
Can developers access the model?
RadixArk says it released BF16, FP8 and NVFP4 checkpoints under the Apache 2.0 license.
Have the benchmark results been independently confirmed?
Not in the supplied material. The figures come from RadixArk’s release; independent testing and fuller evaluation details are still needed to assess them.
What is unclear about the training run?
The release describes training as lasting 18 days, but also refers to weeks and 63,000 GPU-hours. It does not reconcile those accounts in the material provided.
Primary source: Hugging Face · via ThorstenMeyerAI.com
