Ground Truth.
AI, checked against the source.

News · 2026-11-19

DaisyRose-27B targets agent tasks with reinforcement learning

RadixArk released DaisyRose-27B, a downloadable AI model trained on curated agent tasks, and reports that it more than doubled its base model's score on one terminal benchmark. The company also claims substantial training efficiency gains, but the complete comparison and release training data are not public in the reviewed material.

Key facts

DaisyRose builds on Qwen3.8-27B instead of training a new base model from scratch. Its model card lists an Apache-2.0 license, multimodal capabilities and a 262,144-token context. RadixArk also offers FP8 and NVFP4 versions. Disk size and GPU memory are different constraints: weights must be loaded and a running system needs additional space for its working state. The reviewed material does not establish a minimum inference VRAM number.

Reinforcement learning trains a model using feedback on attempted tasks. Here, RadixArk says it probed the base model to identify failures, selected around 1,200 long-running tasks, and checked for reward shortcuts. Think of it as designing a practice course around mistakes instead of simply providing another textbook. The site's lesson on verifiable rewards explains why the scoring mechanism matters.

The company reports the largest relative improvement on Terminal Bench 4.1, with smaller gains on other listed tests. Those are its measurements, not a blind external reproduction. The score describes a particular benchmark and setup; it does not guarantee improvement on every production coding task. The task set and graders are not available in the reviewed materials, so outsiders cannot fully audit how representative or resistant to shortcuts they were.

The systems claim is separate. RadixArk reports 37% versus 16% model-compute utilization in “comparable runs,” from which it calculates 57% less GPU time. Asynchronous task generation and training offer a plausible way to reduce idle hardware, but the detailed matched configurations and internal backend were not published. Even if the relative comparison holds, 144 GPUs over 18 days is a substantial absolute resource commitment. The public Miles framework does not by itself reproduce the release run.

A Hugging Face discussion describes a user conversion and smoke test; a Hacker News discussion includes questions about cost and generalization. Neither verifies the claimed training savings. The important next test is independent evaluation on tasks and serving configurations that reflect real use.

RadixArk says its selection process began with a larger candidate pool, probed the base model's failures and tuned tasks to an appropriate difficulty. It reports that some individual attempts lasted up to 90 minutes and involved over 100 tool calls. This helps explain why training infrastructure matters: the trainer must coordinate long, variable-length work rather than handle only short, uniform exercises.

The run reportedly assigned 128 GPUs to generating agent trajectories and 16 to training. The company's other cost claims include about twice the end-to-end throughput and a lower sandbox-service cost, but these measure different parts of the system and rely on separate company comparisons. They should not be added together or treated as one independently audited saving. A useful outside test would publish the workload, hardware, software versions, quality checks and total cost for both systems.


Primary source, verified: read the paper →

Key questions

What is DaisyRose-27B?

It is RadixArk’s downloadable Qwen3.8-27B derivative, trained with reinforcement learning on curated agent tasks.

How large is the download and how much GPU memory does it need?

The BF16 checkpoint is about 52 GB; the reviewed primary evidence does not state a verified minimum inference VRAM requirement.

Was the claimed GPU-time saving independently reproduced?

No independent reproduction or complete matched protocol was found in the reviewed materials.
Cite this

APA

Ground Truth. (2026, November 19). DaisyRose-27B targets agent tasks with reinforcement learning. Ground Truth. https://groundtruth.day/news/daisyrose-27b-reinforcement-learning-agent-release.html

BibTeX

@misc{groundtruth:daisyrose-27b-reinforcement-learning-agent-release,
  title  = {DaisyRose-27B targets agent tasks with reinforcement learning},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/daisyrose-27b-reinforcement-learning-agent-release.html}
}

Topics: open-weights · reinforcement-learning · agents

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.