RealFeedApp beta

DaisyRose-27B outperformed original Qwen3.8-27B on every benchmark

The Qwen3.8-27B-based model was trained on 1,200 tasks over 18 days using 144 H200 GPUs. It outperforms the original Qwen3.8-27B on all benchmarks.

Field
training language models
What they did
Released the DaisyRose-27B model by training it on 1,200 tasks using reinforcement learning methods in 18 days on 144 GPUs.
Why it matters
This approach increased the pass@1 score on the challenging Terminal-Bench 4.1 benchmark from 5.6% to 12.0%.
#reinforcement learning#qwen#terminal-bench#rlhf#radixark

Read the original →