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

🔍 Read the full analysis: DaisyRose-27B: Rethinking Efficiency After AI Training on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

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 full evaluation details are not provided in the source material.

RadixArk has released DaisyRose-27B, its first model, and says reinforcement learning on long-horizon agent tasks raised its Terminal-Bench 4.1 pass@1 score from 5.6% to 12.0%. The model is based on Qwen3.8-27B; the reported benchmark results come from the company, and independent verification is not included in the available material. For more context, see the original analysis of DaisyRose-27B.

RadixArk says it post-trained the base model using Miles, its reinforcement-learning framework, on a curated set of about 1,200 tasks selected from thousands of candidates developed with Eigent AI. The tasks were designed around errors made by the original model and required connected steps such as planning and tool use. The company says it reviewed attempts and removed or revised tasks that could be passed without solving the intended problem.

According to the release, individual attempts could run for up to 1.5 hours, make more than 100 tool calls and produce more than 256,000 output tokens. Longer training sessions could span up to one million tokens, with harness summarization used to carry work across trajectories. RadixArk says the training run used 144 H200 GPUs and took 18 days. It also describes an asynchronous setup in which 128 GPUs handled agent-task inference and 16 trained on the resulting trajectories.

The company 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 comparison uses the higher score from BF16 and FP8 runs; where both formats were tested, the results differed by less than one point. Checkpoints in BF16, FP8 and NVFP4 are available under the Apache 2.0 license, according to RadixArk.

At a glance
announcementWhen: Released; the source describes an 18-da…
The developmentRadixArk released its first model, DaisyRose-27B, and reported benchmark and training-efficiency results from post-training it on multi-step 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.

Why Long-Horizon Agent Training Matters

The release addresses a practical challenge in AI: models used as agents must often work through multiple steps, tools and extended interactions, rather than answer a single prompt. If the reported gains hold up under independent testing, DaisyRose-27B could offer evidence that reinforcement learning on task sequences can improve performance on this kind of work.

RadixArk also frames the release around the cost of training. It reports 37% model FLOPs utilization for its internal backend, compared with 16% for its open-source trainer on comparable runs, and says the backend used 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; the supplied material does not give enough evaluation detail to establish how broadly they apply.

The model is a post-trained version of an existing foundation model, not a new architecture. Its potential relevance is therefore both practical and methodological: the results may interest teams evaluating open-licensed agent models and training approaches, but benchmark scores alone do not establish reliability or performance in everyday deployments.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

From Qwen3.8-27B to Agent Tasks

RadixArk says DaisyRose-27B retains the base model’s architecture, tokenizer and chat template, as well as image understanding and a 256,000-token context window. The release describes post-training rather than a newly designed model architecture. It points readers to the model card for model properties and inference settings.

The training tasks ran in fresh sandboxes managed by a 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 active at one time. It says 99% of sandbox creations succeeded over nine weeks, with the success rate rising above 99.95% since late October. The company also says the infrastructure cost about 80% less than using popular cloud sandbox providers, though the source does not provide the underlying cost comparison.

The release presents the run as an 18-day training effort using 144 H200 GPUs, but also says an internal-backend run lasted weeks and consumed 63,000 GPU-hours. The material provided does not explain how those descriptions relate.

“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 are RadixArk’s own reports; it includes no independent evaluation. The release summary does not provide enough detail to assess statistical variation, consistency across repeated runs or the full settings used for each benchmark. Although RadixArk says its model card contains comparison sources and evaluation settings, those details are not included here.

It is also unclear how much of the score improvement reflects the reinforcement-learning method, the selection and design of tasks, or other evaluation and model-use choices. The apparent difference between an 18-day training run and a run described as lasting weeks and consuming 63,000 GPU-hours is unresolved. The source material’s account of how sandbox infrastructure costs were reduced is incomplete as well.

Independent Reproduction Is the Next Test

RadixArk says readers can access the Apache 2.0 checkpoints and consult the model card for properties and inference settings. The company recommends serving the model with SGLang and DSpark speculative decoding and says its SGLang cookbook provides launch configurations for different GPUs and checkpoint formats.

The next meaningful test is whether independent evaluators can reproduce the reported benchmark comparisons using clearly documented settings, including checkpoint format and evaluation procedure. Further details on repeated-run variation, the training timeline and the cost methodology would help readers judge the claims. The supplied material does not identify a date for an independent evaluation or another scheduled milestone.

Key Questions

What is DaisyRose-27B?

DaisyRose-27B is RadixArk’s first released model, based on Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks, according to the company.

What benchmark improvement does RadixArk report?

RadixArk reports that Terminal-Bench 4.1 pass@1 rose from 5.6% for the base model to 12.0% for DaisyRose-27B. The supplied material does not include an independent check of that result.

Is DaisyRose-27B available to use?

RadixArk says BF16, FP8 and NVFP4 checkpoints are available under the Apache 2.0 license. The company points users to its model card for properties and inference settings.

Have the benchmark and efficiency claims been independently verified?

No independent evaluation is included in the source material. It also leaves open questions about repeated-run consistency, full benchmark settings and how the training-time descriptions fit together.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

From Concept To Reality: Inkling And The Future Of AI Innovation

Thinking Machines has launched Inkling, a 975-billion-parameter multimodal model for text, image, and audio reasoning, available on Hugging Face with high hardware demands.

Claude, NIST Curve Seeds, And The 1990S Timestamps: What Anthropic Says

Anthropic says Claude linked NIST elliptic-curve seed data to 1990s timestamps. The finding is preliminary, and no cryptographic compromise is confirmed.

Zig Creator Calls Spade a Spade, Anthropic Blows Smoke

Zig creator publicly criticizes Anthropic, accusing them of dishonesty; Anthropic responds with vague statements, escalating the dispute.

Voice Dictation Software for Faster Blogging

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