steampunque/Ministral-3-8B-Instruct-2512-MP-GGUF

Mixed Precision GGUF layer quantization of Ministral-3-8B-Instruct-2512 by mistralai

Original model: https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16

The hybrid quant employs different quantization levels on a per layer basis to increased flexibility of trading off performance vs file size. Less parameter bits are used at deep layers and more bits at cortex layers to simultaneously optimize quantized size and model performance. This quant was optimized for similar size to Q4_K_M quant with improved performance while using all K quants to increase processing efficiency on old GPUs or CPUs.

The Q4_K_H layer quant is as follows:

   LAYER_TYPES='[
   [0 ,"Q4_K_L"],[1 ,"Q4_K_M"],[2 ,"Q4_K_S"],[3 ,"Q4_K_M"],[4 ,"Q4_K_S"],[5 ,"Q4_K_M"],[6 ,"Q4_K_S"],[7 ,"Q4_K_S"],
   [8 ,"Q4_K_S"],[9 ,"Q4_K_S"],[10,"Q4_K_S"],[11,"Q4_K_S"],[12,"Q4_K_S"],[13,"Q4_K_L"],[14,"Q4_K_S"],[15,"Q4_K_S"],
   [16,"Q4_K_M"],[17,"Q4_K_S"],[18,"Q4_K_M"],[19,"Q4_K_S"],[20,"Q4_K_M"],[21,"Q4_K_S"],[22,"Q4_K_M"],[23,"Q4_K_S"],
   [24,"Q4_K_M"],[25,"Q4_K_S"],[26,"Q4_K_M"],[27,"Q4_K_S"],[28,"Q4_K_M"],[29,"Q4_K_L"],[30,"Q5_K_S"],[31,"Q5_K_M"],
   [32,"Q5_K_L"],[33,"Q6_K_S"]
   ]'
   FLAGS="--token-embedding-type Q4_K --output-tensor-type Q6_K --layer-types-high"

This quant was optimized for good reasoning performance across a curated set of test prompts.

Comparison:

Quant	size	PPL	Comment
Q4_K_M	5.2e9	6.6	-
Q4_K_H	5.3e9	6.6	slightly larger than Q4_K_M

Usage:

This is a vision capable model. It can be used together with its multimedia projector layers to process images and text inputs and generate text outputs. The mmproj file is made available in this repository. To test vision mode follow the docs in the mtmd readme in the tools directory of the source tree https://github.com/ggml-org/llama.cpp/blob/master/tools/mtmd/README.md .

Performance:

On a 4070 with cuda backend gen rate is about 77tps with ~37k context available (F16 KV) and ~63k available (Q8_0 KV)

The model is trained at 16k context which can be extended to 256k using YARN:

-rope-scaling yarn --yarn-orig-ctx 16384 --rope_scale 16

For other than 256k context set rope_scale to the fraction of configured context size / 16384.0.

NOTE: For improved inference performance, context size may be limited to 16k. Unless large context is needed, this is recommended for best performance.

Benchmarks:

A full set of vision benchmarks for the model is given here: https://huggingface.co/spaces/steampunque/benchlm

Download the file from below:

Link	Type	Size/e9 B	Notes
Ministral-3-8B-Instruct-2512.Q4_K_H.gguf	Q4_K_H	5.3e9 B	~Q4_K_M size with better performance
Ministral-3-8B-Instruct-2512.mmproj.gguf	F16	0.86e9 B	multimedia projector

A discussion thread about the hybrid layer quant approach can be found here on the llama.cpp git repository:

https://github.com/ggml-org/llama.cpp/discussions/13040