DaisyRose-27B Achieves Major Milestone on Terminal-Bench 4.1 for Long-Horizon Agent Tasks
DaisyRose-27B, a specialized model trained on 1,200 curated long-horizon agent tasks, has demonstrated significant improvements in performance, jumping from 5.6% to 12.0% on the Terminal-Bench 4.1 benchmark. Developed over 18 days of Reinforcement Learning (RL) using 144 H200 GPUs, the model is designed to handle session lengths stretched to 1 million tokens, marking a notable advancement in agentic capabilities.
This performance leap on Terminal-Bench highlights the effectiveness of specialized RL training for complex, long-horizon agentic workflows.
- –Performance more than doubled on the benchmark, suggesting that RL on curated, long-horizon tasks is a critical path for improving agentic reliability.
- –The ability to handle 1M token session lengths is a key technical capability for agents performing multi-step operations over extended periods.
- –Given the training resource requirements (144 H200s for 18 days), this model represents a significant compute investment aimed at high-performance agentic reasoning.
DISCOVERED
17h ago
2026-11-19
PUBLISHED
17h ago
2026-11-19
RELEVANCE
AUTHOR
0x0SojalSec