RadixArk releases DaisyRose-27B, an RL-only Qwen3.8 model
Original title: RadixArk/DaisyRose-27B — new model trending #29 on Hugging Face
Apache-2.0 BF16, FP8 and NVFP4 weights keep the 256K context; Terminal Bench 4.1 pass@1 moves from 5.6 to 12.0.
RadixArk released DaisyRose-27B, its first model, as Apache-2.0 weights in BF16, FP8 and NVFP4. It is a multimodal reasoning model built on Qwen3.8-27B with reinforcement learning only, without supervised fine-tuning or distillation, trained on about 1,200 long-horizon agent tasks for 18 days on 144 H200 GPUs. It keeps the base architecture, tokenizer, chat template, image understanding and 256K context. The card reports Terminal Bench 4.1 pass@1 rising from 5.6 to 12.0.
Unconfirmed: Several comparison scores come from other developers or from Terminal Bench 4.0 when a 4.1 entry is absent.
- Sessions ran up to 1M tokens during the 18-day H200 training run.
- IFBench is 82.4 for DaisyRose-27B versus 79.5 for Qwen3.8-27B.
- Reported DaisyRose scores take the higher of BF16 and FP8 runs.
- Some rival Terminal Bench cells use 4.0 scores when 4.1 is missing.
- Organizations
- RadixArk
- AI models
- DaisyRose-27BQwen3.8-27B
From the original article
7 min read at huggingface.coimage-text-to-text · 0 downloads · 115 likes · license apache-2. 0
By RadixArk · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.