RadixArk released DaisyRose-27B, a post-trained variant of Qwen3.8-27B developed using reinforcement learning on long-horizon agent tasks via the open-source Miles framework. Training consumed 63,000 GPU-hours over 18 days on 144 H200 GPUs, utilizing only 1,200 curated tasks co-developed with Eigent AI. Each task chains interdependent problems requiring up to 1.5 hours, 100+ tool calls, and 256k+ output tokens per attempt, with sessions extending to 1M tokens. Tasks were filtered to target specific model weaknesses, tuned to appropriate solve rates, and hardened against reward hacking. The asynchronous Miles backend achieves 37% MFU versus 16% for the Megatron-based open-source trainer, reducing training compute by 57% and doubling end-to-end throughput to ~16,800 tokens/second across 144 GPUs. Efficiency gains scale to 6.2× on Qwen3.6-35B-A3B and >10× on GLM-5.3-Flash. A robust sandbox service built on AgentENV ran ~944,000 sandboxes (peak 1,566 concurrent) with 99% creation success and >99.95% reliability after late October, at 80% lower cost than cloud providers via dynamic memory allocation. DaisyRose-27B raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% and improves across reasoning, coding, and instruction-following benchmarks. BF16, FP8, and NVFP4 checkpoints are available on Hugging Face.
Read original
hackernews