DaisyRose-27B reports RL-only training and a sharp agent benchmark gain
RadixArk presents DaisyRose-27B as an Apache-2.0 open-weights model trained with RL alone on curated long-horizon agent tasks, with BF16, FP8, and NVFP4 releases and Qwen3.8 compatibility. It reports Terminal-Bench 4.1 pass@1 improving from 5.6% to 12.0% in 18 days; training details, evaluation setup, artifacts, and independent replication remain key checks.
Sources (2)
Updated Nov 18, 2026