- Post-training for long-horizon agents completed with just 1,200 curated tasks on 144 H200 GPUs in 18 days
- Terminal-Bench 4.1 pass@1 improved from 5.6% to 12.0%, with gains across reasoning, instruction-following and coding benchmarks
- BF16 and FP8 checkpoints released under Apache 2.0
AI CURATED HOT STORY · 11/19/2026
RadixArk Open-Sources 27B Agent Model Doubling Terminal Task Success Rate
RadixArk has released DaisyRose-27B, its first open-source agent model built on Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks. The model was trained on only 1,200 curated tasks using 144 H200 GPUs for 18 days, and it doubles the pass@1 on Terminal-Bench 4.1 from 5.6% to 12.0%, while supporting up to 1M token context per session.
Technically, DaisyRose-27B scores 75.5 on Terminal-Bench 2.1 and 46.2 on DeepSWE v1.1, with broad gains across reasoning, instruction-following, and coding benchmarks. RadixArk has open-sourced three weight formats—BF16, FP8, and NVFP4—under the Apache 2.0 license, facilitating deployment and further development across different hardware setups.
The release demonstrates that a high-quality small task set combined with efficient training can significantly enhance an agent's long-horizon capabilities under limited compute, offering the open-source community a reproducible training paradigm and potentially accelerating terminal automation and software engineering agents.
Hourly heat curve
23 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 2 reports · Latest first- Hugging Face BlogT1·72 pts
- Hacker News AIT2·72 pts
- Long-horizon agent RL completed with just 1,200 curated tasks, 18 days and 144 H200 GPUs, supporting sessions up to 1M tokens
- Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, with 75.5 on Terminal-Bench 2.1 and 46.2 on DeepSWE v1.1
- BF16, FP8 and NVFP4 checkpoints released openly for varied hardware deployment and further development