Last updated: November 21, 2026
DaisyRose-27B is an open-weight coding and reasoning model that RadixArk released on November 18, 2026. It is Qwen3.8-27B trained further with reinforcement learning only. It reads text and images, has a 262,144-token context, runs on one GPU, and is available as GGUF files and through OpenRouter.
Contents
- What is new in DaisyRose-27B
- DaisyRose-27B vs. Qwen3.8-27B
- Weights, sizes and hardware
- How to run DaisyRose-27B locally
- Serve it with SGLang
- Use it through OpenRouter
- Settings and prompt patterns
- Limitations
- Frequently asked questions
What is new in DaisyRose-27B
DaisyRose-27B is RadixArk’s first model. According to the RadixArk release post, the company trained it on about 1,200 curated long-horizon agent tasks for 18 days on 144 H200 GPUs, using its open-source reinforcement learning framework, Miles. Some training sessions ran as long as 1 million tokens.
The official model card says the model uses reinforcement learning only, “without any supervised fine-tuning or distillation.” It keeps the base model’s architecture, tokenizer, chat template, image understanding and context length, so RadixArk calls it a drop-in replacement for Qwen3.8-27B in existing inference and post-training stacks.
The headline result is on Terminal Bench 4.1, a long-horizon terminal benchmark: pass@1 rises from 5.6 to 12.0. If you already run the base model, our Qwen3.8-27B GGUF guide covers that model’s local setup.
DaisyRose-27B vs. Qwen3.8-27B
RadixArk reports gains over the base model on all seven benchmarks in its table. The scores below are RadixArk’s own, from the model card, run in thinking mode at reasoning effort xhigh.
| Benchmark | DaisyRose-27B | Qwen3.8-27B | Change |
|---|---|---|---|
| Terminal Bench 4.1 | 12.0 | 5.6 | +6.4 |
| Terminal Bench 2.1 | 75.5 | 73.0 | +2.5 |
| DeepSWE v1.1 | 46.2 | 42.2 | +4.0 |
| SWE Bench Pro v2 | 94.7 | 91.3 | +3.4 |
| HLE (text only) | 34.3 | 33.9 | +0.4 |
| IFBench | 82.4 | 79.5 | +2.9 |
| GPQA Diamond | 90.7 | 89.2 | +1.5 |
Against larger open models the picture is mixed. In the same table, Thinking Machines’ Inkling 2 Small scores 13.0 on Terminal Bench 4.1, 72.1 on DeepSWE v1.1 and 44.1 on HLE, ahead of DaisyRose-27B on all three. See our Inkling 2 Small open weights guide and, for a laptop-sized rival, our Inkling 2 Mini guide.
Weights, sizes and hardware
The model has 27.8B parameters in total, including a 0.46B vision encoder and a 0.42B multi-token-prediction head. It uses 64 decoder layers that mix linear-attention (Gated DeltaNet) and full-attention layers in a 3:1 pattern. Inputs are text and images; output is text.
| Version | Size | Best for |
|---|---|---|
| BF16 (RadixArk/DaisyRose-27B) | About 52 GB | Data-center GPUs such as H200, B200, GB300 |
| FP8 (RadixArk/DaisyRose-27B-FP8) | About 29 GB | Single large GPU serving |
| NVFP4 (RadixArk/DaisyRose-27B-NVFP4) | About 23 GB | Blackwell GPUs only, including RTX 5090 |
| GGUF Q4_K_XL (recommended) | 17.7 GB | Laptops, workstations, consumer GPUs |
| GGUF Q3_K_XL | 13.2 GB | 16 GB GPUs |
| GGUF Q2_K_XL | 10.0 GB | Smallest; reasons longer |
Sizes come from the model card and the GGUF repository. Image input with GGUF also needs the 0.9 GB vision projector file, mmproj-DaisyRose-27B-BF16.gguf. The SGLang cookbook notes that on a 32 GB RTX 5090 only NVFP4 is practical.
How to run DaisyRose-27B locally
The GGUF model card gives commands for Ollama and llama.cpp; LM Studio can load the same files.
- Pick a file that fits your memory: Q4_K_XL (17.7 GB) for most machines, Q3_K_XL (13.2 GB) for a 16 GB GPU.
- With Ollama, run
ollama run hf.co/RadixArk/DaisyRose-27B-GGUF:Q4_K_XL. - With llama.cpp, start an OpenAI-compatible server:
llama-server -hf RadixArk/DaisyRose-27B-GGUF:Q4_K_XL --jinja -ngl 999 -c 65536. The--jinjaflag enables tool calling through the embedded chat template. - Point your client at
http://localhost:8080/v1with model name DaisyRose-27B. - Set sampling to temperature 1.0, top_p 0.95, top_k 20 and min_p 0.
Give each request at least 64K tokens of context, because the default reasoning effort can produce tens of thousands of thinking tokens. For more on local runners, see our LM Studio agent remote control guide.
Serve it with SGLang
For servers, RadixArk “strongly recommends” SGLang with DSpark speculative decoding, using the lmsysorg/sglang:latest Docker image. Key launch flags from the model card and cookbook:
--reasoning-parser qwen3 --tool-call-parser qwen3_coder: without these, clients receive thinking and tool calls as raw text.--speculative-algorithm DSPARK --speculative-draft-model-path RadixArk/Qwen3.8-27B-DSpark: the draft model trained for Qwen3.8-27B carries over.- Alternative:
--speculative-algorithm EAGLEuses the built-in MTP head with no extra download. --context-length 262144for the full window; on RTX PRO 6000, RTX 5090 and DGX Spark add--attention-backend flashinfer.
Use it through OpenRouter
If you would rather not host it, OpenRouter lists the model as radixark/daisyrose-27b. When checked on November 21, 2026, it was served by DeepInfra in BF16 at:
- $0.15 per million input tokens
- $1.875 per million output tokens
- $0.0375 per million cached input tokens
The listing shows a 262,144-token context, up to 235,929 output tokens, tool calling, structured outputs, and reasoning efforts of xhigh (default), medium, low and none. Our Hugging Face OpenRouter routing guide explains provider routing and failover.
Settings and prompt patterns
- Leave max_tokens unset on hard problems; RadixArk warns that long thinking runs to tens of thousands of tokens.
- Keep reasoning in history: in multi-turn chats, send earlier turns back with their
reasoning_content. The chat template keeps it by default. - Lower effort for speed: pass
{"reasoning_effort":"medium"}in chat template kwargs, or disable thinking with{"enable_thinking": false}. - Frame agent work as long jobs: the model was trained on chained, multi-stage tasks with more than 100 tool calls, so give it a clear goal, the repository layout and a test command to verify its work.
Limitations
- Benchmarks are self-reported: the scores come from RadixArk’s own runs, and comparison numbers are as reported by other developers.
- Small gains on knowledge tasks: HLE improves only 0.4 points over the base model.
- Larger models still lead on hard agentic tasks such as DeepSWE v1.1.
- NVFP4 needs Blackwell, and the Q2_K_XL file reasons longer and hits context limits more often.
- Thinking is verbose at the default xhigh effort, which raises cost and latency on paid APIs.
For another large open-weight release this month, see our Hy4 open weights guide.
Frequently asked questions
What is DaisyRose-27B?
DaisyRose-27B is RadixArk’s first model, released November 18, 2026. It is Qwen3.8-27B post-trained with reinforcement learning only, with no supervised fine-tuning or distillation. It is an open-weight thinking model that accepts text and images, has a 262,144-token context, and targets agentic coding and terminal work.
Can I run DaisyRose-27B locally?
Yes. RadixArk publishes GGUF files for llama.cpp, Ollama and LM Studio. The recommended Q4_K_XL file is 17.7 GB, and Q3_K_XL (13.2 GB) is suggested for 16 GB GPUs. In Ollama, run ollama run hf.co/RadixArk/DaisyRose-27B-GGUF:Q4_K_XL.
How much does the DaisyRose-27B API cost?
On OpenRouter, the model ID radixark/daisyrose-27b was listed at $0.15 per million input tokens, $1.875 per million output tokens and $0.0375 per million cached input tokens, served by DeepInfra in BF16, when checked on November 21, 2026. Prices can change.
Is DaisyRose-27B free for commercial use?
The weights are released under the Apache 2.0 license, and the base model, Qwen3.8-27B, is also Apache 2.0. Check the LICENSE file in the Hugging Face repository for the full terms before deploying.
How do I turn off thinking in DaisyRose-27B?
Send chat_template_kwargs with enable_thinking set to false in the request’s extra_body when serving with SGLang or llama.cpp. To keep thinking but shorten it, set reasoning_effort to medium; the default is xhigh.
Sources
- Hugging Face: RadixArk/DaisyRose-27B model card
- Hugging Face: RadixArk/DaisyRose-27B-GGUF
- RadixArk blog: DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency (November 18, 2026)
- SGLang documentation: DaisyRose-27B cookbook
- OpenRouter: RadixArk DaisyRose 27B
Image: DaisyRose-27B logo from the official RadixArk/DaisyRose-27B Hugging Face repository (Apache 2.0), used for editorial purposes. FreeAnswers is not affiliated with RadixArk.