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

🔍 Read the full analysis: How DaisyRose-27B Advances Post-Training Efficiency 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 the supplied materials do not include independent verification or enough detail to reconcile parts of the training account.

RadixArk has released DaisyRose-27B, a Qwen3.8-27B model post-trained with reinforcement learning on multi-step agent tasks, and reports that its Terminal-Bench 4.1 pass@1 score rose from 5.6% to 12.0%. The release includes checkpoints under the Apache 2.0 license; the benchmark results and claims about training efficiency are company-reported and have not been independently verified in the supplied material.

RadixArk says it trained the model using Miles, its reinforcement-learning framework, on a curated set of about 1,200 tasks. The tasks were selected from thousands of candidates developed with Eigent AI and were designed, the company says, around errors made by the original model. Many required connected steps such as planning and tool use, rather than a single short response.

According to the release, an individual attempt could run for up to 1.5 hours, make more than 100 tool calls and generate over 256,000 output tokens. Longer sessions could span as many as one million tokens, with summarization used to carry work across trajectories. RadixArk says it reviewed attempts and removed or revised tasks that could be passed without solving the intended problem.

The company reports improvements 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 each reported score uses the higher result from BF16 and FP8 runs, and that results differed by less than one point where both formats were tested. The release offers BF16, FP8 and NVFP4 checkpoints under Apache 2.0.

At a glance
announcementWhen: Released; the company says the training…
The developmentRadixArk released DaisyRose-27B checkpoints and reported benchmark gains from reinforcement-learning post-training 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.

The Cost of Training Longer Agent Runs

The release addresses a practical problem for AI developers: whether reinforcement learning can improve models on work that involves many steps, tools and extended interaction, not just short question-and-answer exchanges. A higher result on Terminal-Bench 4.1 would be relevant to teams evaluating models for terminal-based tasks, but a benchmark score alone does not establish performance in day-to-day deployment.

RadixArk also frames the project as an efficiency result. It reports 37% model FLOPs utilization on its internal backend, compared with 16% for an open-source trainer on comparable runs, and says its approach 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, against about 8,300 for the open-source trainer using the same number of GPUs. These are company comparisons; the supplied account does not give enough evaluation detail to determine how well they generalize to other workloads or hardware setups.

Openly licensed checkpoints make it possible for researchers and developers to inspect and test the model, subject to the license terms. Reproduction could help establish whether the reported gains hold under consistent benchmark settings and whether the training approach offers similar efficiency beyond RadixArk’s infrastructure.

Products Worth Considering

DJI Mini 4K Drone Combo with 4K Camera, 3-Axis Gimbal, and Two Batteries

DJI Mini 4K Drone Combo with 4K Camera, 3-Axis Gimbal, and Two Batteries

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

As an affiliate, we earn on qualifying purchases.

A Qwen Base Model, Not a New Architecture

DaisyRose-27B is described as post-training of an existing Qwen3.8-27B foundation, rather than a new model architecture. RadixArk says it retained the base model’s architecture, tokenizer, chat template, image understanding and 256,000-token context window. The model card is identified by the company as the source for model properties and inference settings.

The release describes an asynchronous training setup: 128 H200 GPUs handled agent-task inference while 16 GPUs trained on the resulting trajectories, with updated models sent back to inference after each step. RadixArk says the run used 144 H200 GPUs and its headline summary gives a duration of 18 days. Elsewhere, it describes a run on the internal backend lasting weeks and consuming 63,000 GPU-hours; the supplied material does not explain how those accounts relate.

For task execution, the company says it used 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. RadixArk also says 99% of sandbox creations succeeded over nine weeks, with the rate exceeding 99.95% since late October, and claims its infrastructure cost about 80% less than popular cloud sandbox providers.

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

— RadixArk

Benchmark and Training Details Still Open

The benchmark, speed, reliability and cost figures in the supplied account come from RadixArk’s release. No independent evaluation is included, and the material provided does not report statistical variation, run-to-run consistency or full settings for each benchmark. The company says its model card contains comparison sources and evaluation settings, but those details are not reproduced in the source material here.

It is also unclear how much of the benchmark improvement reflects the reinforcement-learning method itself, the curated task set, or other evaluation and model-use choices. The release does not reconcile its description of an 18-day training run with the separate account of training over weeks and 63,000 GPU-hours. Its cost comparison also lacks an explanation of the method used to calculate the claimed savings.

As a result, the reported figures should be read as claims about this training run and evaluation, not as independently established measures of performance or cost across other deployments.

Independent Replication Is the Next Test

RadixArk says the Apache 2.0 checkpoints are available and points users to the model card for properties and inference settings. It recommends serving the model with SGLang and DSpark speculative decoding, and says an SGLang cookbook includes launch configurations for different GPUs and checkpoint formats.

The next useful step is independent testing of the benchmark comparisons using documented settings and the released checkpoints. That work could clarify whether the gains are reproducible, how they vary by checkpoint format, and whether the reported training-efficiency comparisons hold under comparable conditions. The supplied material does not give a timetable for such an evaluation or identify an independent group that has completed one.

Key Questions

What is DaisyRose-27B?

DaisyRose-27B is RadixArk’s Qwen3.8-27B model post-trained with reinforcement learning on long-horizon agent tasks. The company says it retained the base model’s architecture and several other properties.

What benchmark improvement did RadixArk report?

RadixArk reports that pass@1 on Terminal-Bench 4.1 increased from 5.6% to 12.0% compared with the original model. The result is company-reported and has not been independently verified in the supplied material.

Are the model checkpoints available?

Yes. The release offers BF16, FP8 and NVFP4 checkpoints under the Apache 2.0 license, according to RadixArk.

Has the training-efficiency claim been independently confirmed?

No independent confirmation is included in the supplied source. RadixArk reports efficiency comparisons for its internal backend, but the full evaluation settings and independent reproduction are not available here.

How long did training take?

The headline summary says the run took 18 days. The release also describes a run lasting weeks and using 63,000 GPU-hours, without explaining how those descriptions fit together.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

Explosive Speculation Surrounds OpenAI’s Groundbreaking Q* Algorithm

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

Can Genie 4 Understand The Real World? Google Calls It A “Key Stepping Stone” To AGI

Google calls Genie 4 a step toward AGI, but its real-world understanding has not been tested and visual realism alone is not proof.

The Impact Of SenseTime’s Open-Source SenseNova-Vision On AI Development

SenseTime has open-sourced its SenseNova-Vision model, a unified vision system, marking a strategic shift and expanding China’s open-source AI efforts.

Import AI 478: Math For Alignment; ProgramBench; And Mirendil Shows What RSI Means

Import AI 478 covers AI-assisted math formalization, a sharp ProgramBench score change, and OpenAI’s report on agents communicating through cache timing.