知Yizhi Daily
← Back to this edition

AI

RadixArk releases DaisyRose-27B, strengthening Qwen3.8 agent capabilities with 1,200 long-horizon tasks

Edition 2026-11-19 · 08:00 Beijing time

RadixArk has released DaisyRose-27B, its first model based on Qwen3.8-27B, using the Miles reinforcement learning framework and 1,200 selected long-horizon agent tasks, trained on 144 H200 GPUs for 18 days. Terminal-Bench 4.1 pass@1 improved from 5.6% to 12.0%, and BF16, FP8 and NVFP4 checkpoints are available.

RadixArk says DaisyRose-27B is based on Qwen3.8-27B and uses Miles for reinforcement-learning post-training. Training targets long-horizon agent tasks, with sessions up to 1 million tokens; ultimately only 1,200 carefully selected tasks were used, trained on 144 H200 GPUs for 18 days. The model's pass@1 on Terminal-Bench 4.1 rose from the original model's 5.6% to 12.0%, and it also improved on other reasoning, instruction-following and programming benchmarks.

The training tasks were built by RadixArk in collaboration with Eigent AI, with candidate tasks covering fields including machine learning, software engineering and the sciences, and about 1,200 ultimately selected. Each task was redesigned around errors exposed by the original Qwen3.8-27B and calibrated to a solution-rate range that was neither too easy nor too difficult.

The tasks emphasize long chains and prevention of reward gaming: a single task contains multiple interdependent problems, and the agent must go through stages such as reasoning, planning and execution; each attempt runs in a sandbox for up to 1.5 hours, including more than 100 tool calls and over 256,000 output tokens. The team also reviews the model's attempt process and fixes or removes tasks that can be passed without genuinely solving the problem.

Training uses Miles' fully asynchronous mode: 128 H200 GPUs run agent task inference and 16 H200 GPUs run the trainer, with agents continuously working and the model learning synchronously; updated models are pushed to inference GPUs after each step. RadixArk says its backend model FLOPs utilization reached 37%, compared with 16% for a Megatron-based open-source trainer in similar runs, reducing GPU time required for equivalent training by 57% and delivering roughly 2x faster end-to-end speed.

Original sources

Hacker News ↗
Back to this editionOpen interactive reader