You no longer need a giant AI lab to build a better coding bot. Startup RadixArk took a free model from Alibaba's Qwen team, trained it for 18 days on 1,200 hand-picked tasks, and more than doubled its score on one of the hardest tests for AI agents. Then it gave the result away for free.
DaisyRose-27B: Key Takeaways
- RadixArk released DaisyRose-27B, its first model, built on the free Qwen3.8-27B
- Its Terminal-Bench 4.1 score jumped from 5.6% to 12.0% after 18 days of training on 144 Nvidia H200 chips
- It beat the original Qwen model on every benchmark RadixArk tested, from coding to reasoning
- The model is free to use and change under the Apache 2.0 license
RadixArk announced the release on Nov. 18 in a blog post on Hugging Face, the site where AI developers share models. "How much can you improve a small model in 18 days on 144 GPUs?" the company asked on X.
What RadixArk Actually Did
RadixArk did not build a model from scratch, which costs big labs billions. It took Qwen3.8-27B, a 27-billion-parameter model anyone can download, and trained it further with reinforcement learning, where the AI tries a task, gets graded and learns from what worked.
The tasks were the secret. RadixArk and partner Eigent AI built thousands of candidates in machine learning, software engineering and science, then kept about 1,200. Each one targeted a mistake the original model kept making, needed many steps to solve and could only be passed by actually solving it, not by gaming the grader. A single attempt could run up to 1.5 hours, with more than 100 tool calls.
| 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 |
Cheaper Training Is The Real Product
RadixArk is really selling its training tools, and the model is the proof. It trained DaisyRose-27B with Miles, its free reinforcement learning software, and its own training backend, which it says got 37% of each chip's peak power into useful work, against 16% for a popular open-source trainer. That means the same job needs 57% less GPU time, the company says, with even bigger gains on some other model designs.
The test environments matter too. Every attempt runs in a sandbox, a fresh walled-off copy of the task. RadixArk ran about 944,000 of them between mid-September and mid-November, with up to 1,566 at once, and says the cloud bill was about 80% lower than using popular sandbox providers. The whole run used about 63,000 GPU-hours.
Also Read: Developers Are Getting $1B Of AI Training For Free As Meta Floods Hugging Face
Why It Matters
The big AI race is shifting from who builds the largest model to who squeezes the most out of an existing one. If a small team can double a free model's score on a tough agent test in under three weeks, companies with their own data could build strong, specialized bots without frontier-lab budgets.
Keep the numbers in perspective. A 12.0% pass rate on Terminal-Bench 4.1 still means the model fails most of those tasks, and the gains on easier tests were a few points each. The results come from RadixArk's own testing, and the company has asked users to share where the model falls short.
What To Watch Next
Watch whether outside developers can reproduce the gains on their own tasks, and whether other teams adopt Miles to train their own agents. If they do, expect a wave of small, cheap, specialized models built on top of free ones from Qwen, Meta and others, putting pressure on the big labs that charge for their best models.
Read Next: Hugging Face Clones OpenAI’s Math Tests, Admits Verifier Is ‘Early And Rough’
Originally published on Fathom News.
Top comments (0)