DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

RadixArk/

DaisyRose-27B

$0.15

in

$1.875

out

$0.038

cached

/ 1M tokens

$0.20

in

$2.50

out

$0.05

cached

/ 1M tokens

25% off

DaisyRose-27B is RadixArk's multimodal reasoning model, trained from Qwen3.8-27B with reinforcement learning only on long-horizon agent tasks. It improves on the base model in agentic coding, reasoning and instruction following, with thinking that can be enabled or disabled, image understanding and a 256K-token context.

Deploy Private Endpoint
Public
Zero retention
262,144
JSON
Function
Multimodal
ProjectLicense
RadixArk/DaisyRose-27B cover image
api
RadixArk/DaisyRose-27B cover image
DaisyRose 27B

Ask me anything

0.00s

Settings

Model 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.

Highlights

  • Reinforcement learning only: trained from Qwen3.8-27B on long agent tasks of several interdependent problems each, with more than 100 tool calls and more than 256K output tokens per attempt.
  • Agentic coding: Terminal Bench 4.1 12.0 (base 5.6), Terminal Bench 2.1 75.5, DeepSWE v1.1 46.2, SWE Bench Pro v2 94.7.
  • Reasoning and instructions: GPQA Diamond 90.7, HLE (text only) 34.3, IFBench 82.4.
  • Thinking control: thinking is on by default and can be turned off per request; reasoning_effort sets its depth.
  • Vision: image understanding from the base model, unchanged.

Model Overview

PropertyValue
Model typeMultimodal causal language model with a vision encoder; thinking model
ArchitectureQwen3.8-27B: 64 decoder layers in a 3:1 pattern of linear-attention (Gated DeltaNet) and full-attention layers, hidden size 5,120, a 27-layer vision encoder, one multi-token-prediction layer
Parameters27.8B in total, including a 0.46B vision encoder and a 0.42B multi-token-prediction head
Context length262,144 tokens
Input / outputText and images / text
Served on DeepInfraFP8 weights with DSpark speculative decoding, on SGLang

Benchmark Results

DaisyRose-27B against its base model, open models of similar or larger size, and GPT-6 Luna. Scores as reported by RadixArk; the best score in each row is highlighted.

DaisyRose-27BQwen3.8-27BMuse Glimmer-30BQwen3.6-35B-A3BMiMo V2.6 FlashInkling 2 SmallBeam 1.0Muse Spark OpenGPT-6 Luna
Size
Activated parameters27B27B30B3B15B17B23B42B–
Total parameters27B27B30B35B309B594B501B816B–
Agentic coding
Terminal Bench 4.112.05.60.50.016.113.06.46.615.2
Terminal Bench 2.175.573.051.744.987.681.380.176.473.0
DeepSWE v1.146.242.24.40.067.972.143.360.066.6
SWE Bench Pro v294.791.380.473.3–97.7–95.0–
Reasoning
HLE (text only)34.333.922.022.235.144.134.743.938.5
GPQA Diamond90.789.283.586.088.793.090.689.090.5
Instruction following
IFBench82.479.577.064.4––79.6––

All DaisyRose-27B runs use thinking mode at temperature 1.0, top_p 0.95, top_k 20, min_p 0, reasoning effort xhigh, with no max_tokens cap within the 256K context. Terminal Bench 4.1: 8 repetitions per task, mini-swe-agent harness with summarization. Terminal Bench 2.1: 5 repetitions, Terminus 2 harness. DeepSWE v1.1: 8 repetitions, mini-swe-agent. SWE Bench Pro v2: 5 repetitions, mini-swe-agent, full 642-task set. HLE: 5 repetitions, no tools. IFBench: 4 repetitions, prompt-level loose accuracy. GPQA Diamond: 8 repetitions, answer order reshuffled each time. Other models' numbers are as reported by their developers or leaderboards; see the original model card for every source.

Usage Notes

  • Recommended sampling in thinking mode: temperature 1.0, top_p 0.95, top_k 20, min_p 0.
  • The answer comes back in content and the thinking in reasoning_content. In multi-turn chats, send earlier turns back with their reasoning_content.
  • reasoning_effort: low, medium, or high (the model's deepest level, also the default). none turns thinking off.
  • Leave max_tokens unset for hard problems: long thinking can run to tens of thousands of tokens.

Training

DaisyRose-27B is trained from Qwen3.8-27B with reinforcement learning only; the vision encoder and the multi-token-prediction head are the base model's, unchanged. The tasks were built with Eigent AI and each one reviewed together with the model's attempts. Training ran fully asynchronously with Miles, RadixArk's open-source reinforcement-learning framework: 128 H200s ran the agent on its tasks, served by SGLang, while 16 more trained on the finished trajectories, about 63,000 GPU-hours in all.

License

DaisyRose-27B is released under the Apache 2.0 license. It is a derivative of Alibaba's Qwen3.8-27B, which is also licensed under Apache 2.0.