DaisyRose-27B Brings Efficiency Into Focus For AI Post-Training
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: DaisyRose-27B Brings Efficiency Into Focus For AI Post-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 for long-horizon agent tasks. The company reports a Terminal-Bench 4.1 pass@1 increase from 5.6% to 12.0% and says the checkpoints are available under Apache 2.0, but the results have not been independently verified.

RadixArk has released DaisyRose-27B, a 27-billion-parameter model post-trained with reinforcement learning on tasks requiring extended planning and tool use. The company says the model raised Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%; the checkpoints are available under the Apache 2.0 license, but the reported results have not been independently verified.

RadixArk describes DaisyRose-27B as its first model, built on Qwen3.8-27B using the company’s Miles reinforcement-learning framework. The release says the model retains the base model’s architecture, tokenizer, chat template, image understanding and 256,000-token context window. The announcement is therefore about post-training an existing foundation, rather than introducing a new model architecture.

The company says it assembled a curated set of about 1,200 tasks from thousands of candidates developed with Eigent AI. The tasks were selected to address errors made by the original model and were intended to require connected steps, such as planning and tool use. According to RadixArk, an attempt could last up to 1.5 hours, make more than 100 tool calls and generate more than 256,000 output tokens. Longer sessions could span as many as one million tokens, with summarization used to carry work across trajectories.

RadixArk reports higher scores than the base model on all seven listed benchmarks: Terminal-Bench 4.1 and 2.1, DeepSWE v1.1, SWE Bench Pro v2, text-only HLE, IFBench and GPQA Diamond. The company says it selected the higher score from BF16 and FP8 runs for each benchmark; where both formats were tested, the results differed by less than one point. Its release also offers BF16, FP8 and NVFP4 checkpoints under Apache 2.0.

At a glance
announcementWhen: Released; the company says training too…
The developmentRadixArk released its first model, DaisyRose-27B, and reported benchmark gains after reinforcement-learning post-training on extended 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-Running Agents

The release addresses a practical problem for AI developers: an agent may need to maintain a plan, use tools repeatedly and recover from errors over many steps, rather than answer a single prompt. If RadixArk’s results hold up, the reported Terminal-Bench gain could point to a way of improving those extended tasks through post-training. The other benchmark increases, as reported by the company, may indicate that the gains are not limited to one test, but the supplied material does not provide enough detail to judge their size or reliability independently.

RadixArk also presents training efficiency as part of the announcement. It says its internal backend achieved 37% model FLOPs utilization, compared with 16% for its open-source trainer on comparable runs, and describes the difference 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, versus about 8,300 for the open-source trainer on the same number of GPUs. These are company comparisons; their relevance to other teams depends on evaluation conditions, workloads and reproducibility.

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.

How RadixArk Built the Training Run

RadixArk says it ran training asynchronously, assigning 128 H200 GPUs to agent-task inference and 16 GPUs to training on the resulting trajectories. Updated models were sent back to the inference GPUs after each training step. The tasks ran in fresh sandboxes managed by a service based on the open-source AgentENV project. The company reports about 944,000 sandbox runs from mid-September to mid-November, with as many as 1,566 running at one time.

The release says sandbox creation succeeded in 99% of cases over nine weeks, rising above 99.95% since late October. RadixArk also claims its sandbox infrastructure cost about 80% less than popular cloud sandbox providers. The material provided does not describe the cost calculation or explain the method behind that comparison. The company recommends serving the model with SGLang and DSpark speculative decoding and points to an SGLang cookbook for launch configurations.

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

— RadixArk

Benchmark and Training Details to Verify

The benchmark, speed, reliability and cost figures in this article come from RadixArk’s release. The supplied material contains no independent evaluation and does not provide enough information to assess statistical variation, run-to-run consistency or the complete settings for every benchmark. The company says its model card includes comparison sources and evaluation settings, but those details were not included in the source material available here.

It is also unclear how much of the reported improvement reflects the reinforcement-learning approach, the curated task set or other choices in evaluation and model use. The release describes the model as trained in 18 days using 144 H200 GPUs, but also says an internal-backend run lasted weeks and consumed 63,000 GPU-hours. It does not explain how those time descriptions relate. The provided account of the claimed sandbox cost reduction is incomplete, leaving its calculation and scope unclear.

Independent Reproduction of the Scores

RadixArk says readers can inspect the Apache 2.0 checkpoints and consult the model card for model properties and inference settings. The next useful test is whether independent evaluators can reproduce the benchmark comparisons using clearly specified task versions, prompts, sampling settings and checkpoint formats. That would help establish how consistently DaisyRose-27B performs and whether its gains extend beyond the company’s own setup.

Further technical documentation could also clarify the discrepancy between the stated 18-day training period and the description of a run lasting weeks, as well as the basis for the training-speed and sandbox-cost comparisons. Until then, the release establishes that the model and checkpoints are available and records what RadixArk reports—not a third-party confirmation of those claims.

Key Questions

What is DaisyRose-27B?

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

What benchmark improvement did RadixArk report?

RadixArk reports that Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0% compared with the original model. The result has 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. It recommends SGLang and DSpark speculative decoding for serving.

Are the reported training and benchmark results independently confirmed?

No independent evaluation is included in the source material. Details needed to assess repeatability, benchmark settings and the relationship between the stated training durations remain unclear.

Primary source: Hugging Face · via ThorstenMeyerAI.com

You May Also Like

Why Isn’t The Industry Freaking Out About DeepSeek 4.1 Flash?

A developer’s month using DeepSeek 4.1 Flash highlights low costs and cache efficiency, but the account is anecdotal, not independent proof.

Omarchy Foundation Makes A $1.5 Million Deal For SpaceXAI’s Grok

A Techaeris headline reports a $1.5 million Grok deal with the Omarchy Foundation, but the parties, terms and status are unconfirmed.

Could NVIDIA Cosmos 3 Open Models Shape Interactive World Simulation?

NVIDIA has released four open-weight Cosmos 3 Nano checkpoints for driving and robotics simulation. Their usefulness for policy training remains unmeasured.

Announcing Ranveer Singh As Brand Ambassador For Ray-Ban And Ray-Ban Meta In India Along With Exciting New Updates To Our AI Glasses

Meta names Ranveer Singh ambassador for Ray-Ban and Ray-Ban Meta in India, where Gen 3 AI glasses start at INR 44,300.