Exploring Post-Training Efficiency With DaisyRose-27B
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Exploring Post-Training Efficiency With DaisyRose-27B 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 that Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, but the results have not been independently verified in the supplied 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 checkpoints are available under the Apache 2.0 license, giving developers access to the model while leaving the reported performance gains awaiting independent verification, as detailed in the original analysis.

The release describes DaisyRose-27B as a post-trained version of Qwen3.8-27B, built using Miles, RadixArk’s reinforcement-learning framework. The company says the model retains the base model’s architecture, tokenizer, chat template, image understanding and 256,000-token context window. It is therefore a training update to an existing foundation model, not a newly described architecture.

RadixArk says it selected about 1,200 tasks from thousands of candidates developed with Eigent AI. The tasks were intended to require linked steps such as planning and tool use, and were designed around errors made by the original model. According to the release, individual attempts could last 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 between trajectories.

The company reports gains over 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. It says comparisons use the higher result from BF16 and FP8 runs for each benchmark, with differences under one point between formats where both were tested. BF16, FP8 and NVFP4 checkpoints are offered under Apache 2.0.

At a glance
announcementWhen: Announced in the supplied release; the…
The developmentRadixArk released its first model, DaisyRose-27B, with Apache 2.0 checkpoints and company-reported benchmark gains after reinforcement-learning post-training.
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.

The Cost of Training Longer Agent Runs

The release addresses a practical problem for AI developers: training models to complete tasks that involve many steps, repeated tool use and extended interaction, rather than producing a single short answer. RadixArk’s reported Terminal-Bench result is its largest listed gain, while reported increases across the other benchmarks suggest the company sees improvement beyond one test. Those results remain company claims until other evaluators can reproduce them under disclosed, comparable conditions.

RadixArk also presents the project as an efficiency result. It reports that its internal backend reached 37% model FLOPs utilization, compared with 16% for its open-source trainer on comparable runs, and describes this 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. If independently confirmed, the figures could inform decisions about the expense of extended agent training. The supplied release does not provide enough evaluation detail to establish how broadly those comparisons apply.

Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

As an affiliate, we earn on qualifying purchases.

From Qwen3.8 to Agent Training

DaisyRose-27B’s reported gains come from post-training an existing model rather than introducing a new model architecture. RadixArk says its reinforcement-learning process used task attempts as training trajectories, with a review process to remove or revise tasks that could be passed without solving the intended problem. This makes the task design and evaluation procedure relevant to interpreting the final scores.

The company describes a fully asynchronous setup: 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 says the tasks ran in fresh sandboxes managed by a service based on AgentENV. The release reports around 944,000 sandbox runs between mid-September and mid-November, and says sandbox creation succeeded 99% of the time across nine weeks, rising above 99.95% since late October. It also claims the infrastructure cost about 80% less than popular cloud sandbox providers, without enough information in the supplied material to assess that comparison.

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

— RadixArk

Benchmark Details Still Need Review

The supplied release is the source for the benchmark, throughput, reliability and cost figures. It does not include an independent evaluation or enough detail to assess run-to-run variation, statistical uncertainty or the complete settings behind each benchmark. RadixArk says its model card contains comparison sources and evaluation settings, but those details are not included in the material provided here.

The relationship between the training timeline descriptions is also unresolved. The release says the training run took 18 days, but elsewhere describes a run lasting weeks and consuming 63,000 GPU-hours; it does not explain how those accounts fit together. The source material also does not establish how much of the reported benchmark change is attributable to the training method, task selection or other evaluation choices. The method behind the claimed sandbox cost reduction is not fully described.

Model Card and Replication Checks

RadixArk points readers to the model card for model properties and inference settings and recommends serving DaisyRose-27B with SGLang and DSpark speculative decoding. The company says its SGLang cookbook includes launch configurations for different GPUs and checkpoint formats. The checkpoints are available under Apache 2.0, allowing outside teams to examine and test them.

The next useful step is an independent reproduction of the benchmark comparisons using clearly specified checkpoints, evaluation settings and scoring procedures. Further details reconciling the training duration and GPU-hour figures would also help readers assess the efficiency claims. Until those checks are available, the reported gains should be treated as RadixArk’s results, not independently established performance.

Key Questions

What is DaisyRose-27B?

It is RadixArk’s first model, described as Qwen3.8-27B post-trained with reinforcement learning on long-horizon agent tasks. The company says it retains the base model’s architecture and 256,000-token context window.

What benchmark improvement did RadixArk report?

RadixArk says Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%. It also reports higher scores than the base model across six other listed benchmarks. These results have not been independently verified in the supplied material.

Can developers access the model?

RadixArk says BF16, FP8 and NVFP4 checkpoints are available under the Apache 2.0 license. The company also points users to its model card and SGLang cookbook for model and serving information.

How long did training take?

The release summary says the run took 18 days using 144 H200 GPUs. Elsewhere, the material describes a run lasting weeks and consuming 63,000 GPU-hours, without explaining how those descriptions relate.

Have the reported results been independently checked?

No independent evaluation is included in the supplied material. The benchmark and efficiency figures are company-reported; reproduction with disclosed evaluation settings would help establish how well they hold up.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

AI Automation Tools For Small Businesses: A Labor Day sales Guide

Discover how small businesses can harness AI automation tools to save time, cut costs, and grow. Practical tips and real-world examples inside.

Rendering Signature Storm Data In AI: The Zero-Image Breakthrough

A new AI technique visualizes storm evolution without images, using procedural graphics synchronized with scrolling. This advances weather data representation.

Revolutionize Data Presentation With AI And Sheets Canvas

Google introduces Sheets Canvas, a Gemini-powered feature that creates interactive dashboards from spreadsheets via natural language prompts, now rolling out globally.

AI Innovation At Its Early Stage: How To Experiment With SpaceXAI’s Grok Bot

SpaceXAI’s Grok bot is now in early beta, offering limited user access to test its conversational AI capabilities amid fierce industry competition.