· via dev.to (home feed)
RadixArk's DaisyRose-27B lifts free Qwen model's Terminal-Bench 4.1 score from 5.6% to 12%
RadixArk reports DaisyRose-27B, a fine-tune of Qwen3.8-27B, more than doubled its Terminal-Bench 4.1 score after 18 days of reinforcement learning, and released it free under Apache 2.0.

RadixArk releases DaisyRose-27B
Startup RadixArk has published its first model, DaisyRose-27B, a retrained version of Qwen3.8-27B, the freely downloadable 27-billion-parameter model from Alibaba's Qwen team. According to Fathom News, which reported the story on dev.to, the release was announced on Nov. 18 in a Hugging Face blog post, and the weights are available under the Apache 2.0 license, so anyone can use or modify them.
The headline result: after 18 days of training on a cluster of 144 Nvidia H200 GPUs, the model's pass rate on Terminal-Bench 4.1, a demanding test of AI agents working in a terminal environment, rose from 5.6% to 12.0%.
Curated tasks instead of a new base model
RadixArk did not train a foundation model from scratch, an endeavor that costs major labs billions. Instead it applied reinforcement learning, in which a model attempts a task, receives a grade, and updates its behavior based on what worked.
The task design appears to be the key. Working with partner Eigent AI, the team generated thousands of candidate problems spanning machine learning, software engineering and science, then trimmed the set to roughly 1,200. Each retained task aimed at a recurring weakness in the base model, required many steps to complete, and could be passed only by genuinely solving the problem rather than finding loopholes in how it was scored. A single attempt could run as long as 1.5 hours and involve more than 100 tool calls.
Gains on every benchmark tested
RadixArk says the retrained model beat the original Qwen on each benchmark it ran:
| Benchmark | Qwen3.8-27B | DaisyRose-27B |
|---|---|---|
| Terminal-Bench 4.1 | 5.6 | 12.0 |
| Terminal-Bench 2.1 | 73.0 | 75.5 |
| DeepSWE v1.1 | 42.2 | 46.2 |
| IFBench | 79.5 | 82.4 |
| GPQA Diamond | 89.2 | 90.7 |
All of these figures come from RadixArk's own evaluations, and the company has asked users to report where the model still falls short.
The training economics are the real pitch
The model also functions as a demonstration of RadixArk's training infrastructure, which is what the company is actually selling. The run used Miles, its freely available reinforcement learning software, paired with its own training backend. RadixArk claims the backend converted 37% of each GPU's peak capability into useful training work, against 16% for a widely used open-source trainer, a difference it says cuts GPU time for an equivalent job by 57%, with larger gains on some other model designs.
Task attempts ran in isolated sandbox environments. RadixArk executed about 944,000 of them between mid-September and mid-November, with as many as 1,566 running at once, and estimates its sandbox bill came in roughly 80% below mainstream providers. The entire effort consumed about 63,000 GPU-hours.
Why it matters
The competitive frontier is arguably shifting from who can build the largest model to who can extract the most capability from an existing one. If a small team can more than double a free model's score on a hard agentic benchmark in under three weeks, organizations sitting on their own domain data may be able to produce strong, specialized agents without frontier-lab budgets.
The numbers deserve perspective, as Fathom News notes. A 12.0% pass rate on Terminal-Bench 4.1 means the model still fails the large majority of those tasks, and the improvements on the other benchmarks were in the low single digits. The results are self-reported and have not yet been independently verified.
What to watch
Two questions will decide whether this matters beyond a single announcement: whether outside developers can reproduce similar gains on their own tasks, and whether other teams adopt Miles for their own training runs. If both happen, expect a wave of small, cheap, specialized models built on free bases from Qwen, Meta and others, which would put pricing pressure on the large labs that charge for their most capable models.
- #open-source
- #reinforcement-learning
- #ai-agents
- #llms
- #qwen