RadixArk releases DaisyRose-27B, a multimodal reasoning model trained with pure RL on Qwen3.8-27B
Original title: RadixArk/DaisyRose-27B · Hugging Face
RadixArk released DaisyRose-27B, a multimodal reasoning model built on Qwen3.8-27B with reinforcement learning only, without SFT or distillation, under Apache-2.0 in BF16, FP8 and NVFP4. Training used about 1,200 curated long-horizon agent tasks with sessions up to 1M tokens, running 18 days on 144 H200s (about 63,000 GPU-hours). Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0 versus the base model; it keeps the base architecture, tokenizer, 256K context and image understanding as a drop-in replacement.
The model card provides full training details and multi-benchmark comparisons, letting readers gauge how a pure RL approach performs on agentic tasks.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.