GEEK HAUS

Summary and commentary: Geek Haus. Editorial responsibility: Jay Jang, CEO

Back to feed

RadixArk releases DaisyRose-27B, boosting Qwen3.8-27B’s Terminal-Bench 4.1 score to 12%

·Hugging Face Blog
read original ↗

Geek Haus editorial team · Sources and editorial policy

EDITOR BRIEF

RadixArk released DaisyRose-27B, an Apache 2.0 model based on Qwen3.8-27B, trained with reinforcement learning on about 1,200 curated long-horizon tasks. Its Terminal-Bench 4.1 pass@1 score rose from 5.6% to 12.0%, and it scored higher than the base model on all listed benchmarks; the company reports its training backend used 57% less GPU time than comparable runs with its open-source trainer.

COMMENTARY

For teams adapting models to tool-using tasks, DaisyRose-27B offers a downloadable checkpoint and a reported gain on Terminal-Bench, while retaining the base model’s 256K context window. The efficiency figures may make long-horizon post-training more practical, but they come from RadixArk’s internal backend and task setup; independent replication and performance on teams’ own workloads remain open questions.

COMMENTS

Discussion