The official repository only provides GGUF for thinking-2601; there was no GGUF for instruct-2601, so it was converted and quantized directly.

Model details

Base model: kakaocorp/kanana-2-30b-a3b-instruct-2601 (Kakao Corp)
Architecture: DeepseekV3 (MLA + MoE, 30B total / 3B active, 48 layers, 128 experts)
Features: Korean-specialized, tool calling support, non-thinking (instruction-tuned), 32K context
Conversion: llama.cpp convert_hf_to_gguf.py → llama-quantize
Chat template: embedded in GGUF (ChatML <|im_start|> + Qwen-style tool XML)

Available quantizations

File	Method	Size	Notes
`*-Q8_0.gguf`	Q8_0	~30GB	Highest quality
`*-Q6_K.gguf`	Q6_K	~24GB	High quality
`*-Q5_K_M.gguf`	Q5_K_M	~20GB	Balanced
`*-Q4_K_M.gguf`	Q4_K_M	~17GB	Recommended (practical, 4.85 BPW)

Usage

Ollama

# Modelfile
FROM ./kanana-2-30b-a3b-instruct-2601-Q4_K_M.gguf
PARAMETER num_ctx 8192
PARAMETER stop "<|im_end|>"

ollama create kanana2-instruct -f Modelfile
ollama run kanana2-instruct "Hello"

llama.cpp

llama-cli -m kanana-2-30b-a3b-instruct-2601-Q4_K_M.gguf -p "Hello" -c 8192

License

These GGUF files are distributed under the same Kanana License as the original model. Commercial use (API/cloud services, resale, 10M+ MAU, etc.) requires a separate commercial license from Kakao. See the Kanana License for details.

deul/Kanana-2-30b-a3b-instruct-2601-GGUF

Kanana-2-30b-a3b-instruct-2601 GGUF