AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What DaisyRose-27B Reveals About AI Post-Training Efficiency on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

RadixArk released DaisyRose-27B, a Qwen3.8-27B model post-trained with reinforcement learning on extended agent tasks. The company reports Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, but the supplied results have not been independently verified and some training details are unresolved.

RadixArk has released DaisyRose-27B, a Qwen3.8-27B model post-trained with reinforcement learning on long-horizon agent tasks, and reports that its Terminal-Bench 4.1 pass@1 score rose from 5.6% to 12.0%, as described in the original analysis. The checkpoints are available under the Apache 2.0 license, but the benchmark and training-efficiency figures come from the company and have not been independently verified in the supplied material.

RadixArk says it used its Miles reinforcement-learning framework to train the model on a curated set of about 1,200 tasks, selected from thousands of candidates developed with Eigent AI. The company says tasks were chosen to address errors made by the base model and required connected steps such as planning and tool use. It reports that individual attempts could last up to 1.5 hours, use more than 100 tool calls and produce over 256,000 output tokens.

The release reports higher scores than the original model across seven listed benchmarks: Terminal-Bench 4.1 and 2.1, DeepSWE v1.1, SWE Bench Pro v2, text-only HLE, IFBench and GPQA Diamond. RadixArk says it selected the higher result from BF16 and FP8 runs for each benchmark; where both formats were tested, the scores differed by less than one point. The supplied material gives the Terminal-Bench 4.1 change but does not provide the other benchmark scores.

RadixArk describes an asynchronous training setup using 144 H200 GPUs: 128 handled agent-task inference and 16 trained on the resulting trajectories. It reports 37% model FLOPs utilization on its internal backend, compared with 16% for its open-source trainer on comparable runs. The company also reports about 16,800 tokens per second across 144 GPUs when agent-task time is included, versus about 8,300 for the open-source trainer on the same GPU count.

At a glance
reportWhen: Release timing is not specified in the…
The developmentRadixArk released DaisyRose-27B checkpoints and reported benchmark gains from reinforcement-learning post-training on long-horizon agent tasks.
AI Model Release · Briefing

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.

5.6% → 12.0%
Terminal-Bench 4.1 pass@1
~1,200
Curated agent tasks
144
H200 GPUs used
18 days
Headline training duration

What changed

The release describes post-training an existing model foundation, not a new architecture.

Foundation

Built on Qwen3.8-27B

RadixArk says it retained the architecture, tokenizer, chat template, image understanding and 256,000-token context window.

Training

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.

1
Curate tasks

About 1,200 tasks were selected from thousands of candidates and refined around errors made by the original model.

2
Run agents in sandboxes

An attempt could last up to 1.5 hours, make more than 100 tool calls and produce more than 256,000 output tokens.

3
Train on trajectories

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.

Model FLOPs utilization37%
Open-source trainer utilization16%
Tokens per second across 144 GPUs~16,800
Open-source trainer, same GPU count~8,300

What readers can check

The release offers checkpoints and points to the model card for properties and inference settings.

Available
  • BF16, FP8 and NVFP4 checkpoints
  • Apache 2.0 license
  • SGLang and DSpark serving recommendations
Still needed
  • Independent benchmark reproduction
  • Repeated-run and evaluation details
  • Reconciliation of 18 days with weeks-long training and 63,000 GPU-hours
CaveatBenchmark, speed, reliability and cost figures come from RadixArk; the supplied material does not include an independent evaluation.
The takeaway
DaisyRose-27B offers an open route to test reported gains on long-running agent tasks, while independent reproduction is needed to establish how broadly they apply.

What the Agent Benchmark Gain Shows

The release addresses a practical obstacle in building AI agents: performance on a task may depend on sustained planning, tool use and recovery across many steps, not just producing a strong answer in one exchange. If the reported benchmark results hold under independent testing, DaisyRose-27B would offer evidence that reinforcement-learning post-training can improve performance on extended agent tasks without changing the underlying model architecture.

The reported training figures also speak to the cost of producing such improvements. Higher utilization and throughput could matter to teams deciding whether to train models on lengthy, interactive tasks. But the comparisons depend on the particular workloads, hardware and measurement methods used. The release does not establish that the efficiency gap will apply across other models or training systems, nor that benchmark gains will carry over to routine deployments.

The distinction between a model release and independently validated results matters for developers considering adoption. Apache 2.0 checkpoints make the model available for inspection and testing, while reproducible evaluations would help users judge its capabilities and costs against alternatives. For now, the figures are best treated as company-reported results, not a settled measure of general agent performance.

Products Worth Considering

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD — Midnight

As an affiliate, we earn on qualifying purchases.

A Post-Training Run on Agent Tasks

DaisyRose-27B is described as a post-trained version of Qwen3.8-27B, rather than a new model architecture. RadixArk says it retains the base model’s tokenizer, chat template, image understanding and 256,000-token context window. The release offers BF16, FP8 and NVFP4 checkpoints under Apache 2.0.

RadixArk says task attempts ran in fresh sandboxes managed by a service based on the open-source AgentENV project. It reports about 944,000 sandbox runs from mid-September to mid-November, with as many as 1,566 running at once. The company says 99% of sandbox creations succeeded over nine weeks, with the rate above 99.95% since late October, and claims its infrastructure costs about 80% less than popular cloud sandbox providers. These figures have not been independently checked in the supplied material.

The release characterizes the work as an 18-day training run using 144 H200 GPUs. Elsewhere, it says a run on the internal backend lasted weeks and consumed 63,000 GPU-hours. The source material does not explain whether these descriptions refer to different periods or measurements.

““DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.””

— RadixArk

Questions Around the Reported Results

The benchmark, throughput, reliability and cost figures in the supplied material come from RadixArk’s release. It does not provide an independent evaluation or enough information to assess statistical variation, consistency across repeated runs, or every benchmark’s full settings. The company says its model card includes evaluation settings and comparison sources, but those details were not included in the source material provided here.

It is also not clear how much of the reported improvement is attributable to the reinforcement-learning method, the task selection, or other choices in evaluation and model use. RadixArk’s descriptions of an 18-day run and a run lasting weeks with 63,000 GPU-hours are not reconciled. The supplied description of the claimed sandbox cost savings is incomplete, leaving the calculation and comparison basis unclear.

These gaps do not establish that the results are inaccurate, but they limit what readers can conclude from the release alone. In particular, benchmark gains do not by themselves show how the model will perform across different tools, tasks or real-world operating conditions.

Independent Testing and Model Access

RadixArk has made the checkpoints available under Apache 2.0 and points readers to a 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 useful step is independent reproduction of the benchmark comparisons, with the task definitions, evaluation settings and checkpoint formats clearly documented. Further detail explaining the 18-day and multiweek training descriptions would also help readers assess the claimed compute efficiency. Until those checks and clarifications are available, the model’s published scores and infrastructure comparisons remain company-reported.

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 says it retains the base model’s main architecture and listed features.

What benchmark gain did RadixArk report?

RadixArk reports that Terminal-Bench 4.1 pass@1 increased from 5.6% to 12.0%. The supplied material does not include independent verification of that result.

How was the model trained?

The company says it used its Miles framework and about 1,200 curated agent tasks. Its reported setup used 128 H200 GPUs for task inference and 16 for training, in an asynchronous process.

Can developers access the checkpoints?

RadixArk says BF16, FP8 and NVFP4 checkpoints are available under the Apache 2.0 license. It also points users to a model card and serving guidance.

What remains unverified?

The benchmark and efficiency figures have not been independently confirmed in the supplied material. Evaluation details, run-to-run variation, the basis for some cost comparisons and the difference between the stated 18-day and multiweek training periods remain unclear.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

Google Beam’s Expansion Brings New Regions And Business Partners

Google says Beam is shipping in six countries, supported by 18 channel partners, with bookable rooms at select Industrious sites from October.

AI Software For Small Business Automation: Compare The Options

A comparison of Zapier and Make finds a trade-off: easier setup and broad app connections versus more control over complex AI workflows.

Inside Holo4: Building More Versatile Computer-Use Agents

H Company says Holo4 combines GUI, code and tool use in one model. Its OSWorld scores await independent verification.

A Hardware Checklist For Teams Planning AI Automation In 2026

A planning guide reviews workstation, storage, power, audio and switching hardware while warning teams to verify specifications and compatibility.