DNA3.0-27B-GGUF

DNA3.0-27B GGUF

Unofficial community GGUF quantization of dnotitia/DNA3.0-27B, intended for local inference with LM Studio and llama.cpp.

The upstream model is a 27B dense vision-language model based on Qwen/Qwen3.6-27B, with additional Korean, enterprise persona, and uncensored post-training by Dnotitia. This repository does not add a separate abliteration pass; it only converts and quantizes the published upstream weights.

Files

For text-only use, download the main model. For image input, also download the mmproj file and select it as the vision projector if your runtime does not detect it automatically.

LM Studio

1. In LM Studio, search for kuper0201/DNA3.0-27B-GGUF.
2. Download DNA3.0-27B-Q4_K_M.gguf.
3. Download mmproj-DNA3.0-27B-Q8_0.gguf as well if you need image input.
4. Load the model with a modest context size first, then increase it only if memory permits.

The Q4 model is about 15.4 GiB before runtime overhead. A 16 GB unified-memory Mac may use heavy swap and run slowly; 24 GB or more memory is recommended. The upstream model supports very long contexts, but local memory use grows with the selected context length.

Conversion details

• Source: dnotitia/DNA3.0-27B
• Converter/runtime: llama.cpp commit 6ea215d
• Intermediate precision: BF16
• Final quantization: Q4_K_M
• Vision projector quantization: Q8_0
• The conversion used --no-mtp because the published model contains 64 main transformer blocks plus separate MTP tensors.

SHA-256

745c63bdb76bada3f5b525eeab73bad9a53988e7eb98be865ba103e9ee41b084  DNA3.0-27B-Q4_K_M.gguf
28b0e39206f481da17c9ba1aa3d4d41f9df23d203bc51bb24f3f7e5c63e2cfe8  mmproj-DNA3.0-27B-Q8_0.gguf

License and responsible use

The upstream model is published under the Apache License 2.0. Please review the upstream model card for its limitations, biases, and responsible-use guidance. Quantization can also introduce quality differences relative to the original BF16 weights.