RadixArk open-sources DaisyRose-27B: 18 days on 144 H200s, Terminal-Bench 5.6% to 12.0%

BanghuaZ · x · 2026-11-19

At SGLang Summit, RadixArk launched DaisyRose-27B, its first open-weight model: RL training on Qwen3-27B with 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs. Using Miles and its in-house training backend, the team stretched agent sessions up to 1M tokens and lifted Terminal-Bench 4.1 from 5.6% to 12.0%. The Day 0 pin shown was a summit Easter egg.

Original post →

More from coding & agent

coding & agent channel →