Vibitely Daily

Thursday, 19 November 2026, story 4 of 7

RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned with RL

Checkpoints come in BF16, FP8 and NVFP4; RadixArk says Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%.

Illustration: A large stone block with a few deliberately chiselled notches, each notch aligned with a small brass gear beside it, while a rougher, unchiselled block stands behind.

What happened

RadixArk has released DaisyRose-27B, its first model. It is built on Qwen3.8-27B and trained with reinforcement learning using Miles, the company's open-source framework. RadixArk says it is releasing BF16, FP8 and NVFP4 checkpoints, available on Hugging Face.

The company says Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, and that the model also gains on other reasoning, instruction-following and coding benchmarks. Training used about 1,200 tasks built with Eigent AI and ran 18 days on 144 H200 GPUs, with sessions up to 1M tokens. The scores are RadixArk's own, from its internal harness.

What it means

The most reusable part is the method. RadixArk says it ran the base model on every candidate task, rewrote tasks to hit its mistakes, tuned solve rates, and removed any task that could be passed without real work. Teams fine-tuning agents on small budgets can borrow that discipline without any RadixArk software.

The efficiency numbers need care. The 57% cut in GPU time compares RadixArk's internal backend with the open-source trainer. The post does not say the backend or its sandbox service is released, so Miles users may not see those gains. Licence terms and the task set are also not stated. A pass rate of 12.0% is more than double the base model's but still low in absolute terms, and the company is also offering to work with teams on post-training. Independent runs on real workloads will say more than the post does.

What to do: Test the checkpoints on your own agent tasks before relying on them; the training speed-ups come from a backend RadixArk has not said it is releasing.

Sources

Filed under Open source. Published .

"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.