DaisyRose-27B GGUF

GGUF conversion of RadixArk/DaisyRose-27B, for use with llama.cpp.

DaisyRose-27B is a 27B-parameter multimodal reasoning model based on Qwen3.8-27B. See the original model card for training details, evaluations, and limitations.

Files

File Type Purpose
DaisyRose-27B-Q4_K_M.gguf Q4_K_M Main language model, including its built-in multi-token-prediction layer
mmproj-DaisyRose-27B-F16.gguf F16 Vision encoder/projector for image inputs

Use a recent llama.cpp version with Qwen3.5 support. The model has a native 262,144-token context; choose a smaller context if your memory is limited.

Text inference

llama-server -hf angt/DaisyRose-27B-GGUF:Q4_K_M \
  -c 32768 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0

Then open http://localhost:8080 or use the OpenAI-compatible /v1/chat/completions endpoint.

For terminal use:

llama-cli -hf angt/DaisyRose-27B-GGUF:Q4_K_M \
  -c 32768 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0

Image inference

Download both files, then pass the projector to the server:

hf download angt/DaisyRose-27B-GGUF \
  DaisyRose-27B-Q4_K_M.gguf mmproj-DaisyRose-27B-F16.gguf \
  --local-dir DaisyRose-27B-GGUF

llama-server \
  -m DaisyRose-27B-GGUF/DaisyRose-27B-Q4_K_M.gguf \
  --mmproj DaisyRose-27B-GGUF/mmproj-DaisyRose-27B-F16.gguf \
  -c 32768 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0

The main GGUF also works without mmproj for text-only inference.

Conversion details

The official BF16 repository was used as the conversion source instead of dequantizing the FP8 repository, avoiding an additional loss of precision. Both repositories contain the same DaisyRose-27B fine-tune.

  • Source: RadixArk/DaisyRose-27B
  • Source revision: 293a7a2ab54aaea78098ee917c546390667521dc
  • Requested FP8 counterpart: RadixArk/DaisyRose-27B-FP8
  • Converter: llama.cpp commit 78e05298f0d68abff41d08d79fa7469264b2aff0
  • Quantization: standard Q4_K_M, without an imatrix
  • Chat template, tokenizer, generation defaults, and MTP tensors are preserved by the converter
  • Validation: text generation answered an arithmetic prompt correctly; multimodal inference identified a test image correctly

Checksums

SHA-256 checksums are listed in SHA256SUMS. Verify a download with:

sha256sum -c SHA256SUMS

Equivalent conversion commands:

python convert_hf_to_gguf.py model-src \
  --outfile DaisyRose-27B-BF16.gguf --outtype bf16 --model-name DaisyRose-27B

llama-quantize DaisyRose-27B-BF16.gguf \
  DaisyRose-27B-Q4_K_M.gguf Q4_K_M

python convert_hf_to_gguf.py model-src \
  --outfile DaisyRose-27B-F16.gguf --outtype f16 \
  --model-name DaisyRose-27B --mmproj

License

Apache-2.0, inherited from the source model. The original license is included in LICENSE.

Downloads last month
351
GGUF
Model size
27B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for angt/DaisyRose-27B-GGUF

Base model

Qwen/Qwen3.8-27B
Quantized
(28)
this model