ymcki/Kimi-Linear-48B-A3B-Instruct-GGUF

This is a repo for experimental GGUFs for the backend agnostic implementation of the Kimi-Linear model support that requires a llama.cpp from this repo. You can git clone it and compile locally.

git clone https://github.com/ymcki/llama.cpp --branch Kimi-Linear
cd llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j 6

If you have enough VRAM, you can run it purely on your graphics card:

./build/bin/llama-cli -m ~/Kimi-Linear-48B-A3B-Instruct-GGUF/Kimi-Linear-48B-A3B-Instruct.Q4_K_M.gguf -c 8192 -ngl 100 --mmap

Otherwise, you can only load the shared experts and KV cache to your graphics card and the rest to CPU and RAM.

./build/bin/llama-cli -m ~/Kimi-Linear-48B-A3B-Instruct-GGUF/Kimi-Linear-48B-A3B-Instruct.Q4_K_M.gguf -c 8192 -cmoe -ngl 100 --mmap

I am going to only make ggufs without imatrix and ggufs with an imatrix based on c4_en_ja_imatrix.txt for better Japanese performance as bartowski and unsloth will make ggufs with English imatrix anyway.

Base perplexity for f16 gguf is 7.291970 ± 0.048577.

Seems like MLA KV cache can only be run at F16 probably due to itself being a kind of compression. You can use this table to see how much context you can run with a single 24GB card.

Quant Type	imatrix	File Size	Delta Perplexity	KL Divergence	Description
Q5_K_M	c4_en_ja_imatrix.txt	34.87GB	7.115874 ± 0.047587	0.074066 ± 0.001537	Good
Q5_K_M	None	34.87GB	7.133672 ± 0.047741	0.074684 ± 0.001535	Good. Slightly worse than imatrix
Q4_K_M	c4_en_ja_imatrix.txt	29.70GB	7.147482 ± 0.047851	0.081894 ± 0.001521	Good. Can run 128k context on a single 32GB card.
Q4_K_M	None	29.70GB	7.172188 ± 0.048107	0.083700 ± 0.00152	Good. Slightly worse than imatrix
MXFP4_MOE	None	27.21GB	7.179840 ± 0.047966	0.088789 ± 0.001544	Good. Can run 240k context on a single 32GB card.
MXFP4_MOE	c4_en_ja_imatrix.txt	27.21GB	7.179840 ± 0.047966	0.088789 ± 0.001544	Good. Same as the no imatrix version.
IQ4_XS	c4_en_ja_imatrix.txt	26.27GB	7.208724 ± 0.048490	0.088246 ± 0.001528	Good. Can run 304k context on a single 32GB card.
IQ4_NL	c4_en_ja_imatrix.txt	27.79GB	7.209342 ± 0.048412	0.087678 ± 0.001532	Doesn't make sense compare to MXFP4_MOE
IQ3_M	c4_en_ja_imatrix.txt	21.55GB	7.368516 ± 0.048425	0.113435 ± 0.001457	Quite Good. Can run 96k context on a single 24GB card.
IQ3_S	c4_en_ja_imatrix.txt	21.33GB	7.448991 ± 0.049167	0.119987 ± 0.001466	Quite Good. Can run 112k context on a single 24GB card.
IQ3_XS	c4_en_ja_imatrix.txt	20.17GB	7.534649 ± 0.049461	0.129645 ± 0.001448	Quite Good. Can run 176k context on a single 24GB card.
Q3_K_S	c4_en_ja_imatrix.txt	21.33GB	7.557247 ± 0.051236	0.131708 ± 0.001521	Quite Good. Can run 112k context on a single 24GB card.
Q3_K_S	None	21.33GB	7.632887 ± 0.051792	0.146355 ± 0.001534	Quite Good but worse than imatrix. Good for CPU use.
IQ3_XXS	c4_en_ja_imatrix.txt	18.99GB	7.780732 ± 0.052592	0.164925 ± 0.001537	Not so good but can run 240k context on a single 24GB card.
IQ2_M	c4_en_ja_imatrix.txt	16.13GB	8.207663 ± 0.054957	0.224437 ± 0.001536	Slightly batter than Q2_K but you can run 400k context on a single 24GB card.
Q2_K	c4_en_ja_imatrix.txt	18.03GB	8.295144 ± 0.057566	0.221437 ± 0.001617	So-so but you can run 288k context on a single 24GB card. Good for performance evaluation.
Q2_K	None	18.03GB	8.648201 ± 0.059234	0.267082 ± 0.001659	Worse than imatrix

As expected, imatrix has no effect on MXFP4_MOE. From this reddit thread, its perplexity is about the same as IQ4_XS but about 6% bigger file size. Here, its perplexity is better than IQ4_XS. This makes it a viable option.