AIsaAIsaStudio 中文 Open AIsa Studio

RadixArk releases DaisyRose-27B, an RL-only multimodal reasoning model built on Qwen3.8-27B

Original title: RadixArk/DaisyRose-27B (image-text-to-text)

RadixArk released DaisyRose-27B, a multimodal reasoning model fine-tuned from Qwen3.8-27B via reinforcement learning only, without SFT or distillation, trained on about 1,200 curated long-horizon agent tasks with sessions of up to 1M tokens for 18 days on 144 H200 GPUs. Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0, and the model keeps the base model's architecture, tokenizer, 256K-token context and image understanding, released under Apache-2.0 in BF16, FP8, and NVFP4.

Why it matters

The model card includes benchmark comparisons against other open models and evaluation details, letting readers gauge how pure RL fine-tuning performs on long-horizon agent tasks.

Read the original ↗

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.

Original article

Hugging Face 热门模型第 13 名:RadixArk/DaisyRose-27B,192 赞,近期下载 166 次。 标签:transformers, safetensors, qwen3_5, image-text-to-text, daisyrose, qwen3.8, reinforcement-learning, conversational, base_model:Qwen/Qwen3.8-27B, base_model:finetune:Qwen/Qwen3.8-27B, license:apache-2.0, endpoints_compatible

README:

DaisyRose-27B

BF16 | FP8 | NVFP4 | GGUF | DSpark draft | Base model | Blog

This repository holds the BF16 weights of DaisyRose-27B.

1. General Information

DaisyRose-27B is RadixArk's first model: a multimodal reasoning model built on Qwen3.8-27B with reinforcement learning only, without any supervised fine-tuning or distillation. It was trained on about 1,200 curated long-horizon agent tasks, with sessions of up to 1M tokens, for 18 days on 144 H200 GPUs. Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0, and the model also improves on most of the other coding, reasoning and instruction-following benchmarks below.

DaisyRose-27B keeps the base model's architecture, tokenizer and chat template, along with its image understanding and 256K-token context, so it is a drop-in replacement for Qwen3.8-27B in existing inference and post-training stacks. Weights are released under Apache-2.0 in BF16, FP8, and NVFP4.

2. Evaluations

DaisyRose-27B against its base model, open models of similar or larger size, and GPT-6 Luna

Benchmark DaisyRose-27B Qwen3.8-27B Muse Glimmer-30B Qwen3.6-35B-A3B MiMo V2.6 Flash Inkling 2 Small Beam 1.0 Muse Spark Open GPT-6 Luna
Activated parameters 27B 27B 30B 3B 15B 17B 23B 42B –
Total parameters 27B 27B 30B 35B 309B 594B 501B 816B –
Terminal Bench 4.1 12.0 5.6 0.5 0.0 16.1 13.0 6.4 6.6 15.2
Terminal Bench 2.1 75.5 73.0 51.7 44.9 87.6 81.3 80.1 76.4 73.0
DeepSWE v1.1 46.2 42.2 4.4 0.0 67.9 72.1 43.3 60.0 66.6
SWE Bench Pro v2 94.7 91.3 80.4 73.3 – 97.7 – 95.0 –
HLE (text only) 34.3 33.9 22.0 22.2 35.1 44.1 34.7 43.9 38.5
IFBench 82.4 79.5 77.0 64.4 – – 79.6 – –
GPQA Diamond 90.7 89.2 83.5 86.0 88.7 93.0 90.6 89.0 90.5

How we evaluated

Terminal Bench 4.1: Averaged over 8 repetitions per task, internal mini-swe-agent harness with summarization enabled.
Terminal Bench 2.1: Averaged over 5 repetitions per task, Terminus 2 harness.
DeepSWE v1.1: Averaged over 8 repetitions per task, mini-swe-agent harness.
SWE Bench Pro v2: Averaged over 5 repetitions per task, mini-swe-agent harness. Qwen3.8-27B score from Table 3 of the SWE Bench Pro v2 paper (full 642-task set, mini-swe-agent).
HLE (text only): Averaged over 5 repetitions per problem, GPT-6 Luna grader, no tools.
IFBench: Averaged over 4 repetitions per prompt (300 prompts), official allenai/IFBench scorer, prompt-level loose accuracy on the final response (thinking excluded).
GPQA Diamond: Averaged over 8 repetitions per question (198 questions), simple-evals multiple-choice template and answer extraction, answer order reshuffled each repetition.

All DaisyRose-27B runs use the base model's recommended thinking-mode settings: temperature 1.0, top_p 0.95, top_k 20, min_p 0, reasoning effort xhigh (the template default), no max_tokens cap within the 256K context. DaisyRose-27B scores are the higher of our BF16 and FP8 runs, which differ by less than one point on every benchmark where we ran both. Other models' numbers are as reported by their developers: see the sources below.

Sources

Terminal Bench 4.1: MiMo V2.6 Flash, Inkling 2 Small and Beam 1.0 from the official leaderboard (mini-swe-agent); GPT-6 Luna from the same leaderboard (Codex, max effort). The other models have no 4.1 entry, so Terminal Bench 4.0 scores are shown instead: Muse Spark Open from Vals.ai (average of 3, mini-swe-agent); Qwen3.8-27B (xhigh), Muse Glimmer-30B and Qwen3.6-35B-A3B from Artificial Analysis (Qwen3.8, Glimmer, Qwen3.6).

Terminal Bench 2.1: model cards for Qwen3.8-27B, Muse Glimmer-30B (Terminus 2), Inkling 2 Small (Terminus 2), Beam 1.0 and Muse Spark Open (Terminus 2). GPT-6 Luna from the Inkling 2 Small model card (Terminus 2). Qwen3.6-35B-A3B from Artificial Analysis (Terminus 2). MiMo V2.6 Flash from the MiMo-V2.6 release (harness not stated).

The original belongs to Hugging Face 热门模型 and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.

Other coverage 2