eaddario/Mistral-Small-3.2-24B-Instruct-2506-pruned-GGUF

Experimental layer-wise + pruned (layers 4 and 5) quantization of mistralai/Mistral-Small-3.2-24B-Instruct-2506

Using LLaMA C++ release b5890 for quantization.

Original model: mistralai/Mistral-Small-3.2-24B-Instruct-2506

From the original model creators:

Mistral-Small-3.2-24B-Instruct-2506 is a minor update of Mistral-Small-3.1-24B-Instruct-2503.

Small-3.2 improves in the following categories:

• Instruction following: Small-3.2 is better at following precise instructions
• Repetition errors: Small-3.2 produces less infinite generations or repetitive answers
• Function calling: Small-3.2's function calling template is more robust (see here and examples)

In all other categories Small-3.2 should match or slightly improve compared to Mistral-Small-3.1-24B-Instruct-2503.

PLEASE READ THIS BEFORE USING THESE EXPERIMENTAL VERSIONS!

An area of personal interest is finding ways to optimize the inference performance of LLMs when deployed in resource-constrained environments like commodity hardware, desktops, laptops, mobiles, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but my focus has been primarily on quantization and pruning.

The method used to produce these experimental versions is covered in Squeezing Tensor Bits: the quest for smaller LLMs, but at a high level it involves using a custom version of llama-imatrix to identify influential tensors, quantize the most important layers to higher bit precision and the less important to lower bits, and remove (prune) one or more layers. This process was partly inspired by Dumitru's et al Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels, and Xin Men's et al ShortGPT: Layers in Large Language Models are More Redundant Than You Expect.

As of version b5125, llama-quantize can perform tensor-wide quantization (TWQ), whereby user-defined tensors are quantized at a specific level, or perform layer-wise quantization (LWQ) by selecting different quantization types per tensor/layer. For example, --tensor-type attn_v=q6_k will quantize all Attention Value tensors at q6_k (TWQ), and --tensor-type "\.([0-9]|1[01257]|31)\.attn_k=q4_k" will quantize Attention Key tensors on layers 0 to 9, 10, 11, 12, 15, 17 and 31 at q4_k, leaving the remaining layers at their default value (LWQ).

As of version b5740, llama-quantize can also prune models during quantisation by providing a comma-separated list in the --prune-layers command line option. The pruning operation will renumber remaining layers to avoid gaps in the sequence, update the relevant model metadata and, if an imatrix is available, it will use the correct importance score vector. This option can be used alongside --tensor-type to perform tensor/layer-wise quantization on selected tensor types, whilst at the same time pruning others. For example:

llama-quantize --tensor-type attn=q6_k --prune-layers 3,7,11 --imatrix imatrix.dat model-f32.gguf model-q4_k_m.gguf q4_k_m

An enhanced version of llama-imatrix generates useful statistics to guide the tensor and layer selection process. --show-statistics will display:

• Σ(Act²): the sum of all squared activations over the tensor (i.e. the Importance Scores)
• Min & Max: minimum and maximum squared activation values
• μ & σ: activations' mean and standard deviation
• % Active: proportion of elements whose average squared activation exceeds a very small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
• N: number of squared activations in the tensor
• Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement)
• E (norm): Normalized entropy.
• ZD Score: z-score distribution as described in 3.1 Layer Importance Scores in the Layer-Wise Quantization paper
• CosSim: cosine similarity between same type tensors with respect to the previous layer (i.e. blk.7.attn_k and blk.6.attn_k)

Please note that statistics are calculated for each individual tensor and should be used to compare between tensors of the same type only. For example, assuming that attn_k in layer 10 has a higher influence during inference than attn_k in layer 7 because its Σ(Act²) is larger makes sense, whilst concluding the same between attn_k and ffn_down does not.

There’s a pull request to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified version will be available on GitHub.

For testing and comparison I use models produced by Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below) but if they don't provide versions of the required model, all tests and comparisons are done against naive quantizations obtained by simply running llama-quantize with no further optimization.

All experimental versions were generated using an appropriate imatrix created from calibration datasets available at eaddario/imatrix-calibration. At its core, an Importance Matrix (imatrix) is a table or, more broadly, a structured representation that scores the relative importance of different features or parameters in a machine learning model. It essentially quantifies the "impact" each feature has on a specific outcome, prediction, or relationship being modelled, and it helps to counterbalance the negative effects of quantization and pruning.

The process to generate these models is roughly as follows:

1. Convert the original model's tensors to GGUF F16*
2. Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
3. Generate an imatrix from selected calibration datasets
4. Determine tensor and layer Importance Score contribution using the enhanced version of llama-imatrix
5. Select an appropriate quant level for each tensor and quantize/prune the model using llama-quantize. In this model's case, layers 4 and 5 have been pruned
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep versions with the best scores
8. Repeat until all desired quants are created. I find that quantizations below Q3/IQ3 are not fit for my purposes and therefore do not usually generate them, but happy to provide other quants on request.

*BF16 would be preferred, but Apple's GPUs don't support it yet, and therefore any operations are executed in the CPU, making it unacceptably slow. This is expected to change in the near term but until then, if you are using Apple kit avoid using any models tagged BF16

Models

Sizes (in GB)

Model	Bartowski	Unsloth	Repo	Shrinkage
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_M	10.7	N/A	10.3	3.7%
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_S	9.9	N/A	9.6	3.1%
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ4_NL	13.5	13.5	12.5	7.4%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_L	12.4	N/A	10.9	12.1%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_M	11.5	11.5	10.0	13.0%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_S	10.4	10.4	9.0	13.5%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	14.3	14.3	12.5	12.6%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_S	13.5	13.5	11.6	14.1%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_M	16.8	16.8	15.3	8.9%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_S	16.3	16.3	14.5	11.0%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q6_K	19.3	19.3	18.9	2.1%
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q8_0	25.1	25.1	20.3	19.1%

Bits per Weight, Perplexity and KL Divergence scores

Model	BPW	μPPL	𝜌PPL	μKLD	RMS Δp
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_M	3.6735	25.645736 ±0.214748	95.06%	0.286497 ±0.002130	12.354 ±0.058
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_S	3.4052	32.512225 ±0.267969	91.81%	0.484943 ±0.002308	17.076 ±0.062
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ4_NL	4.4435	25.450410 ±0.199570	96.78%	0.173444 ±0.001426	11.165 ±0.051
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_L	3.8705	24.405174 ±0.195423	95.33%	0.248285 ±0.001640	12.606 ±0.056
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_M	3.5467	24.750046 ±0.198446	94.92%	0.273478 ±0.001716	13.182 ±0.057
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_S	3.2060	25.982356 ±0.208609	92.33%	0.410016 ±0.002211	15.977 ±0.062
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	4.4492	21.411111 ±0.170166	98.21%	0.092796 ±0.001164	7.244 ±0.048
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-bartowski	4.8620	5.162781 ±0.029306	99.42%	0.022338 ±0.000204	4.694 ±0.046
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-unsloth	4.8620	5.163345 ±0.029336	99.43%	0.022240 ±0.000199	4.648 ±0.045
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_S	4.1390	21.835641 ±0.173204	97.63%	0.126128 ±0.001376	8.429 ±0.051
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_M	5.4555	20.724451 ±0.164266	99.25%	0.035047 ±0.000830	4.125 ±0.051
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_S	5.1466	20.733047 ±0.164062	99.11%	0.042653 ±0.000923	4.585 ±0.050
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q6_K	6.7205	20.450270 ±0.161987	99.58%	0.016946 ±0.000664	2.585 ±0.062
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q8_0	7.2324	20.469495 ±0.162108	99.63%	0.014037 ±0.000644	2.279 ±0.073
Mistral-Small-3.2-24B-Instruct-2506-pruned-F16	16.0003	20.457257 ±0.161275	100%	N/A	N/A

ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores

Scores generated using llama-perplexity with 750 tasks per test, and a context size of 768 tokens.

For the test data used in the generation of these scores, follow the appropiate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_M	65.3333 ±1.7389	80.00	39.6000 ±1.7870	35.0667 ±1.7436	70.9333 ±1.6591	58.19
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ3_S	62.2667 ±1.7711	78.00	39.4667 ±1.7860	36.1333 ±1.7553	72.9333 ±1.6235	57.76
Mistral-Small-3.2-24B-Instruct-2506-pruned-IQ4_NL	62.4000 ±1.7699	79.47	41.4667 ±1.8002	37.4667 ±1.7686	71.3333 ±1.6523	58.43
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_L	64.0000 ±1.7539	77.20	41.4667 ±1.8002	36.8000 ±1.7621	72.8000 ±1.6260	58.45
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_M	64.4000 ±1.7496	77.60	40.8000 ±1.7958	36.9333 ±1.7635	72.9333 ±1.6235	58.53
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q3_K_S	61.0667 ±1.7816	77.87	40.9333 ±1.7967	38.2667 ±1.7759	72.0000 ±1.6406	58.03
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	62.0000 ±1.7736	79.33	41.7333 ±1.8018	39.2000 ±1.7838	71.4667 ±1.6500	58.75
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-bartowski	67.0667 ±1.7172	83.60	45.7333 ±1.8203	36.2667 ±1.7567	79.4667 ±1.4760	62.43
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-unsloth	66.6667 ±1.7225	84.00	45.8667 ±1.8207	35.7333 ±1.7510	79.3333 ±1.4795	62.32
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_S	62.9333 ±1.7648	78.40	40.4000 ±1.7930	38.5333 ±1.7783	71.0667 ±1.6569	58.27
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_M	62.5333 ±1.7686	79.46	41.3333 ±1.7993	38.1333 ±1.7748	72.5333 ±1.6309	58.80
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q5_K_S	63.0667 ±1.7635	79.46	40.9333 ±1.7967	38.5333 ±1.7783	72.2667 ±1.6358	58.85
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q6_K	62.6667 ±1.7674	78.80	41.0667 ±1.7976	37.4667 ±1.7686	72.1333 ±1.6382	58.43
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q8_0	62.2667 ±1.7711	79.60	41.0667 ±1.7976	37.4667 ±1.7686	72.4000 ±1.6334	58.56
Mistral-Small-3.2-24B-Instruct-2506-pruned-F16	61.7333 ±1.7759	79.73	40.8000 ±1.7958	37.4667 ±1.7686	72.4000 ±1.6334	58.43

Tokens per Second - Benchmarks

Scores generated using llama-bench. Naive (llama-quantize with no optimization) Q4_K_M quantization included for comparison.

model	size	params	backend	threads	test	t/s
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	pp512	263.08 ±7.40
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	tg128	28.03 ±0.32
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M	11.63 GiB	22.46 B	Metal,BLAS	12	pp1024+tg1024	47.14 ±0.37
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp512	247.65 ±20.10
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	tg128	27.68 ±1.07
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp1024+tg1024	45.66 ±0.10
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	pp512	253.28 ±16.91
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	tg128	27.75 ±0.79
Mistral-Small-3.2-24B-Instruct-2506-pruned-Q4_K_M-unsloth	13.34 GiB	23.57 B	Metal,BLAS	12	pp1024+tg1024	45.62 ±0.14

Metrics used

Perplexity: one of the key metrics used in NLP evaluation. It measures the quality of a language model by evaluating how well it predicts the next token given a particular sequence of words. A PPL of 1 indicates an exact match between predicted and actual, whereas values greater than one indicate a degree of "surprise" the generated token differs from the expected.

Kullback–Leibler (KL) Divergence: a statistical measure of how much a probability distribution differs from another. When quantizing models (or altering the original tensors in any way for that matter), the closest we can preserve the weights' probability distribution to the original model the better, thus the closest to 0 the better.

AI2 Reasoning Challenge (ARC): a benchmark to evaluate the ability of AI models to answer complex science questions that require logical reasoning beyond pattern matching.

HellaSwag: the Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations (bit of a mouthful!) is a benchmark designed to test commonsense natural language inference. It requires the model to predict the most likely ending of a sentence.

MMLU: the Massive Multitask Language Understanding evaluates LLMs’ general knowledge and problem-solving abilities across 57 subjects, including elementary mathematics, US history, computer science, and law.

Truthful QA: evaluates how well LLMs generate truthful responses to questions. It identifies whether AI models can avoid generating false or misleading information, particularly in areas where human knowledge is prone to misconceptions.

Winogrande: based on the Winograd Schema Challenge, is a natural language understanding task requiring models to resolve ambiguities in sentences involving pronoun references.

Credits

LLaMa C++ has a large and vibrant community of contributors (~1,200 last time I checked) that actively maintain and extend its functionality, adding new models and architectures almost as fast as they appear (considering the breakneck speed at which the AI/ML field is advancing, this alone is a remarkable feat!), and whilst I'm grateful to each and everyone of them, I want to recognise three people in particular: Thank You! Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries, and Iwan Kawrakow for being one of the key authors behind the many quantisation algorithms and the imatrix functionality.