steampunque/Mistral-Small-3.1-24B-Instruct-2503-Hybrid-GGUF

Llama.cpp hybrid layer quantization of Mistral-Small-3.1-24B-Instruct-2503 by mistralai

Original model: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503

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. These quants were specifically optimized so the vision mode of the model produced good outputs with no nonsense words across all the quants on a test case, while reducing the file size significantly to enable full offload (in non vision mode) of the smallest two quants on a 12G VRAM GPU. Three quants are available : Q2_K_H, Q3_K_H, and Q4_K_H. The layer quants are as follows:

Q2_K_H:
LAYER_TYPES='[
   [0 ,"Q2_K"  ],[1 ,"Q2_K_S"],[2 ,"Q2_K"  ],[3 ,"Q2_K_S"],[4 ,"Q2_K"  ],[5 ,"Q2_K_S"],[6 ,"Q2_K"  ],[7 ,"Q2_K_S"],
   [8 ,"Q3_K_S"],[9 ,"Q2_K"  ],[10,"Q3_K_S"],[11,"Q2_K"  ],[12,"Q3_K_S"],[13,"Q2_K"  ],[14,"Q3_K_S"],[15,"Q2_K"  ],
   [16,"Q3_K_S"],[17,"Q2_K"  ],[18,"Q3_K_S"],[19,"Q2_K"  ],[20,"Q3_K_S"],[21,"Q2_K"  ],[22,"Q3_K_S"],[23,"Q2_K"  ],
   [24,"Q3_K_S"],[25,"Q3_K_S"],[26,"Q3_K_S"],[27,"Q3_K_S"],[28,"Q3_K_S"],[29,"Q3_K_S"],[30,"Q3_K_S"],[31,"Q3_K_S"],
   [32,"Q3_K_S"],[33,"Q3_K_S"],[34,"Q3_K_S"],[35,"Q3_K_S"],[36,"Q3_K_M"],[37,"Q3_K_M"],[38,"Q3_K_M"],[39,"Q3_K_M"]
   ]'
FLAGS="--token-embedding-type Q3_K --output-tensor-type Q5_K"

Q3_K_H:
LAYER_TYPES='[
   [0 ,"Q3_K_S"],[1 ,"Q2_K"  ],[2 ,"Q3_K_S"],[3 ,"Q2_K"  ],[4 ,"Q3_K_S"],[5 ,"Q2_K"  ],[6 ,"Q3_K_S"],[7 ,"Q2_K"  ],
   [8 ,"Q3_K_S"],[9 ,"Q2_K"  ],[10,"Q3_K_S"],[11,"Q2_K"  ],[12,"Q3_K_S"],[13,"Q2_K"  ],[14,"Q3_K_S"],[15,"Q2_K"  ],
   [16,"Q3_K_S"],[17,"Q3_K_S"],[18,"Q3_K_S"],[19,"Q3_K_S"],[20,"Q3_K_S"],[21,"Q3_K_S"],[22,"Q3_K_S"],[23,"Q3_K_S"],
   [24,"Q3_K_S"],[25,"Q3_K_S"],[26,"Q3_K_S"],[27,"Q3_K_S"],[28,"Q3_K_S"],[29,"Q3_K_S"],[30,"Q3_K_S"],[31,"Q3_K_M"],
   [32,"Q3_K_M"],[33,"Q3_K_M"],[34,"Q3_K_M"],[35,"Q3_K_M"],[36,"Q3_K_M"],[37,"Q3_K_M"],[38,"Q3_K_L"],[39,"Q4_K_S"]
   ]'
   FLAGS="--token-embedding-type Q4_K --output-tensor-type Q6_K"
fi

Q4_K_H:
LAYER_TYPES='[
   [0 ,"Q3_K_M"],[1 ,"Q3_K_M"],[2 ,"Q3_K_M"],[3 ,"Q3_K_M"],[4 ,"Q3_K_M"],[5 ,"Q3_K_M"],[6 ,"Q3_K_M"],[7 ,"Q3_K_M"],
   [8 ,"Q3_K_M"],[9 ,"Q3_K_M"],[10,"Q3_K_M"],[11,"Q3_K_M"],[12,"Q3_K_M"],[13,"Q3_K_M"],[14,"Q3_K_M"],[15,"Q3_K_M"],
   [16,"Q3_K_L"],[17,"Q3_K_M"],[18,"Q3_K_L"],[19,"Q3_K_M"],[20,"Q3_K_L"],[21,"Q3_K_M"],[22,"Q3_K_L"],[23,"Q3_K_M"],
   [24,"Q3_K_L"],[25,"Q3_K_M"],[26,"Q3_K_L"],[27,"Q3_K_M"],[28,"Q3_K_L"],[29,"Q3_K_M"],[30,"Q3_K_L"],[31,"Q3_K_M"],
   [32,"Q3_K_L"],[33,"Q4_K_S"],[34,"Q4_K_S"],[35,"Q4_K_S"],[36,"Q4_K_M"],[37,"Q5_K_S"],[38,"Q5_K_M"],[39,"Q6_K"]
   ]'
   FLAGS="--token-embedding-type Q4_K --output-tensor-type Q6_K"
fi

These quants were optimized for both good reasoning and vision performance.

Comparison:

Quant	size	PPL	Comment
Q2_K	8.89e9	6.62	not tested, most likely unusable
Q2_K_H	9.8e9	5.96	optimized for good performance in vision mode
Q3_K_H	10.5e9	5.82	slighly better than Q2_K_H
Q3_K_M	11.5e9	5.58	not tested, should work well
Q4_K_H	12.5e9	5.49	slightly smaller than IQ4_XS, similar performance
IQ4_XS	12.9e9	5.38	not tested, should work well

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 . Use of the best available model (Q4_K_H) is recommended to maximize the accuracy of vision mode. To run it on a 12G VRAM GPU use --ngl 32. Generation speed is still quite good with partial offload.

Benchmarks:

A full set of benchmarks for the model will eventually be given here: https://huggingface.co/spaces/steampunque/benchlm

Mistral-Small-3.1-24B-Instruct-2503 compares most closely with gemma-3-27B-it available here: https://huggingface.co/steampunque/gemma-3-27b-it-Hybrid-GGUF . A short summary of some key evals comparing the two models is given here for convenience:

model	gemma-3-27b-it	Mistral-Small-3.1-24B-Instruct-2503
quant	Q4_K_H	Q4_K_H
alignment	strict	permissive
TEST
Winogrande	0.748	0.784
Lambada	0.742	0.798
Hellaswag	0.802	0.899
BoolQ	0.701	0.646
Jeopardy	0.830	0.740
GSM8K	0.964	0.940
Apple	0.850	0.820
Humaneval	0.890	0.853

Download the file from below:

Link	Type	Size/e9 B	Notes
Mistral-Small-3.1-24B-Instruct-2503.Q2_K_H.gguf	Q2_K_H	9.8e9 B	good quality
Mistral-Small-3.1-24B-Instruct-2503.Q3_K_H.gguf	Q3_K_H	10.5e9 B	solid quality
Mistral-Small-3.1-24B-Instruct-2503.Q4_K_H.gguf	Q4_K_H	12.5e9 B	best quality
Mistral-Small-3.1-24B-Instruct-2503.mmproj.gguf	mmproj	0.88e9 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