0

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin(huggingface.co)
RadixArk has released DaisyRose-27B, a new model built upon Qwen3.8-27B that dramatically improves performance on complex agent tasks. The model more than doubles the pass rate on the Terminal-Bench 4.1 benchmark, jumping from 5.6% to 12.0%, while also showing gains across other reasoning and coding benchmarks. This was achieved by training on just 1,200 highly curated tasks using an efficient reinforcement learning framework that completed the process in only 18 days. The success is attributed to high-quality training data, a cost-effective training backend, and a robust sandbox environment for running agent tasks. DaisyRose-27B retains its base model's architecture and 256K context window, with checkpoints available under an Apache 2.0 license.
0 points•by ogg•2 hours ago

Comments (0)

No comments yet. Be the first to comment!

Have an account? Log in to join the discussion.