RadixArk has released DaisyRose-27B, an open-weight model it says was trained with a more efficient post-training setup. The company reports a Terminal-Bench pass@1 score rising from 5.6% to 12.0%, while using 57% less GPU time than its comparable open-source training run.
Watch Desk analysis
What happened
DaisyRose-27B is built on Qwen3.8-27B and trained with reinforcement learning using RadixArk’s Miles framework. RadixArk says it trained the model on 1,200 curated tasks over 18 days, using 144 H200 GPUs and long agent sessions of up to one million tokens. The checkpoints are available under Apache 2.0.
The company attributes the efficiency gains to its internal training backend. It says that backend achieved 37% model FLOPs utilisation, compared with 16% for the open-source trainer on comparable runs. RadixArk also reports roughly twice the end-to-end training throughput, including time spent on agent tasks. These are the company’s figures, not an independent benchmark.
Why it matters
The headline result is not just a higher score on one benchmark. RadixArk is arguing that a more efficient training system can make long-running agent training cheaper to repeat, while releasing the model weights for others to test and use.
The reported Terminal-Bench gain is substantial in relative terms, but the score remains specific to that test. RadixArk says DaisyRose also improves on other reasoning, instruction-following and coding benchmarks; the practical value will depend on how it performs on users’ own work.
Our read
This is a useful release because it puts both a model and a concrete training-efficiency claim on the table. The Apache 2.0 checkpoints make the claim testable by other teams, which is more interesting than another victory lap over a leaderboard. The sensible next move is to try the model on your own tasks and compare the compute bill, not to assume one benchmark settles the matter.
What to watch
- Whether independent evaluations reproduce the reported benchmark gains.
- How DaisyRose-27B performs on practical agent tasks beyond Terminal-Bench.
- Whether other teams can reproduce the training-efficiency results using Miles and the released checkpoints.
Discussion spark: Would you prioritise a model’s benchmark gains or the amount of compute needed to train it when deciding whether it represents real progress?
Sources and evidence
- DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency (18 November 2026, 19:54 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.