The release preserves Kimi-K2.6's native multimodal architecture and is intended as the canonical llama.cpp-compatible GGUF ladder for the original model.

Quick Benchmarks

Check	Original Kimi-K2.6	Kimi-K2.6 GGUF
Official 25-prompt refusal check	Pending	Pending
PPL / KLD reference drift	Pending	Pending

Methodology & Model Notes

Kimi-K2.6 is a large sparse-MoE vision-language model in the Kimi K2 family, exposed through the KimiK25ForConditionalGeneration wrapper with a DeepSeek V3-style text stack.

This release is built directly from moonshotai/Kimi-K2.6 and is intended to provide a clean original-model GGUF ladder without altering the base refusal behavior.

Quant Benchmarks

Quant	Official 25-prompt refusal check	Perplexity	KL divergence
Q8_0	Pending	Pending	Pending
Q6_K	Pending	Pending	Pending
Q4_K_M	Pending	Pending	Pending
Q2_K	Pending	Pending	Pending

Files

Kimi-K2.6-Q8_0/: highest-fidelity quant
Kimi-K2.6-Q6_K/: near-lossless practical quant
Kimi-K2.6-Q4_K_M/: smaller general-use quant
Kimi-K2.6-Q2_K/: lowest standard quant in this ladder
mmproj-Kimi-K2.6.gguf: matching multimodal projector file for llama.cpp vision use

Running

bash

llama-server \
  -m <quant-file.gguf> \
  --mmproj <mmproj-file.gguf> \
  -ngl 999 -c 32768 --jinja -fa

Model Architecture

Spec	Value
Architecture Wrapper	`KimiK25ForConditionalGeneration`
Text Family	DeepSeek V3-style sparse MoE
Text Layers	61
Hidden Size	7168
Experts	384 routed, 8 active per token
Modality	Vision-language
Base Model	moonshotai/Kimi-K2.6

Disclaimer

This is the original Kimi-K2.6 model converted to GGUF. It is not an abliterated release.

Credits

Base model: moonshotai/Kimi-K2.6
GGUF runtime and quantization: llama.cpp

License

This release inherits the base Kimi-K2.6 license.

Modified MIT License.

Youssofal/Kimi-K2.6-GGUF

Kimi-K2.6-GGUF