eaddario/Ministral-3-14B-Reasoning-2512-GGUF

Experimental global target bits‑per‑weight quantization of mistralai/Ministral-3-14B-Reasoning-2512

Using non-standard (forked) LLaMA C++ release b7560 for quantization.

Original model: mistralai/Ministral-3-14B-Reasoning-2512

From the original model creators:

Ministral 3 14B Reasoning 2512

The largest model in the Ministral 3 family, Ministral 3 14B offers frontier capabilities and performance comparable to its larger Mistral Small 3.2 24B counterpart. A powerful and efficient language model with vision capabilities.

This model is the reasoning post-trained version, trained for reasoning tasks, making it ideal for math, coding and stem related use cases.

The Ministral 3 family is designed for edge deployment, capable of running on a wide range of hardware. Ministral 3 14B can even be deployed locally, capable of fitting in 32GB of VRAM in BF16, and less than 24GB of RAM/VRAM when quantized.

Learn more in our blog post and paper.

⚠️ 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 to produce these experimental versions involves using a custom version of llama-imatrix to generate an imatrix that includes the mean activations, and a custom version of llama-quantize, which computes a per-tensor weighted mean squared quantization error and a bias/projection term, to automatically select the lowest error quantization recipe that achieves a global target bits‑per‑weight (bpw). More details on the implementation and test results here

There are two pull requests (#14891 & #15550) to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified versions will be available on GitHub.

For testing and comparison, I use models produced by Bartowski (see credits below) and Unsloth (Daniel and Michael Han do some really interesting stuff!) but when they don't provide versions of the required model, tests and comparisons are against standard quantization obtained by simply running llama-quantize with no further optimizations.

All experimental versions were generated using an appropriate imatrix created from datasets available at eaddario/imatrix-calibration. In llama.cpp, an imatrix is a calibration file derived from running representative text through the model and collecting activation statistics. It is used to weight quantization error so that error in more “important” directions (as estimated from activations) is penalized more heavily.

The process to generate these models is roughly as follows:

1. Convert the original model's safetensors 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 the most appropriate calibration dataset
4. Quantize the baseline model targeting a bpw average, allocating more bits to tensors estimated to matter more (e.g. llama-quantize --target-bpw 4.5678 --keep-bpw-state --imatrix imatrix.gguf baseline-model-F16.gguf 12)
5. Quantize the baseline model targeting a bpw average, treating each tensor equally instead of prioritizing some (e.g. llama-quantize --target-bpw 4.5678 --no-importance --keep-bpw-state --imatrix imatrix.gguf baseline-model-F16.gguf 12)
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep version with the best 𝜌PPL scores (i.e. highest Cor(ln(PPL(Q)), ln(PPL(base))))
8. Repeat until all desired quants are created

*BF16 would be preferred, but F16 performs better on Apple's GPUs

Advantages and disadvantages of the global target bits‑per‑weight quantization process

Advantages

1. Target arbitrary size models

   • When specifying --target-bpw 4.5678 for instance, the algorithm will produce a model (nearly) exactly of that size, which is very useful for maximizing VRAM usage. In a system with 24GB VRAM and a 70B model, standard quants might produce a 16.8GB file (too small, quality left on table) or a 24.1GB file (won't fit). This approach can generate a 23.85GB file to utilize the hardware fully.

2. Data-driven mixed precision often can improve quality at fixed size

   • Instead of using hardcoded heuristics (e.g. make attn_v Q5_K for a 70B model), that may be sub‑optimal for a given architecture or size, the quantization mix is determined by the actual error sensitivity of the specific model's weights. This, in practice, often yields a better quality/size trade-off, especially in aggressive quantization scenarios (1.5 to 3.5 bpw), or for unusual architectures.

   • Please note: llama.cpp’s heuristics have been tuned across many models and are highly optimized; although the target bpw method produces better quality often (>75% based on tests with 130 models from 11 different families), it can also lose in surprising cases.

3. Allows better like-for-like comparisons between models and families

   • Standard llama.cpp quantization uses hardcoded rules like: "use Q4_K_M, except bump some tensors up/down, except fall back if incompatible, except keep some tensors unquantized..." and for that reason, two different models quantized with the same Q4_K_M type can end up with very different bpw (e.g. 4.75 and 4.30).

   • All things being equal, the performance of a model is usually proportional to its overall bpw size; models with a higher bpw tend to perform better than lower bpw models. Since model A has simply been given more bits, it will typically perform better (lower perplexity, better eval scores, etc.) even if the underlying quantization method is identical. That makes comparing the performance not a controlled experiment, because the comparison is between models with different effective compression ratios.

   • --target-bpw tries to address that by making the experiment more controlled: each model gets quantized to land on (approximately) the same global byte budget, so that the models' performance differences are more attributable to architecture/training differences, quantization error behaviour at the same compression ratio, optimizer’s allocation decisions, etc.

Disadvantages

1. Quantization process is significantly slower than standard

   • This approach can take 5x-10x longer as it quantizes a sample of most tensors into 15 different formats, dequantizes them back to floats, computes error diffs, and selects the best size/error option that fits the global bpw budget.

   • However, the --keep-bpw-state option will save the above-mentioned computations to disk so that future quantizations, in the permissible bpw range for the same model, can be generated at normal speed. It also allows to interrupt the computation process and resume it at a later time.

2. The optimization target is only a proxy for the model's performance quality

   • The process minimizes a per-tensor estimated error computed from sampled rows, not actual perplexity or divergence of output distributions (a future version may address this). Since errors interact nonlinearly across layers, there are no guarantees it will select the best possible quantization recipe subject to the bpw size constraint.

   • Furthermore, the process can operate in two modes: giving priority to important tensors (default) or treating each tensor equally (setting the --no-importance option). To my knowledge, there is no computationally feasible way to determine ahead of time which modality will yield better results, and two runs per model may be needed to obtain the best quality, but the default mode usually wins.

3. An imatrix with activations data is required for best results

   • Activation data is required to compute the bias factor (i.e. the systematic error projected onto activation directions). If the imatrix file does not contain activation data, the quantization recipe will likely be sub-optimal.

---

Models

Bits per weight, size, perplexity and KL Divergence scores

Model	BPW	Size (GB)	μPPL	𝜌PPL	μKLD	Same Top-P
Ministral-3-14B-Reasoning-2512-F16	16.0005	27.0	6.436297 ±0.039434	100%	N/A	N/A
Ministral-3-14B-Reasoning-2512-IQ1_L	1.7499	2.96	22.132043 ±0.152503	77.56%	1.208592 ±0.003084	55.222 ±0.128
Ministral-3-14B-Reasoning-2512-IQ2_S	2.2500	3.81	9.588696 ±0.062743	91.12%	0.389679 ±0.001571	72.215 ±0.115
Ministral-3-14B-Reasoning-2512-IQ2_XS	2.1250	3.60	10.290454 ±0.068161	89.60%	0.459912 ±0.001780	70.526 ±0.117
Ministral-3-14B-Reasoning-2512-IQ2_XXS	2.0000	3.38	12.026650 ±0.081067	86.72%	0.607188 ±0.002170	67.300 ±0.120
Ministral-3-14B-Reasoning-2512-IQ3_XXS	2.9999	5.07	7.282053 ±0.045407	97.21%	0.124056 ±0.000618	84.285 ±0.093
Ministral-3-14B-Reasoning-2512-Q2_K	2.4999	4.23	8.565386 ±0.055711	93.92%	0.277481 ±0.001145	76.209 ±0.109
Ministral-3-14B-Reasoning-2512-Q3_K_L	3.7499	6.34	6.761373 ±0.040460	99.00%	0.051564 ±0.000209	89.399 ±0.079
Ministral-3-14B-Reasoning-2512-Q3_K_S	3.2500	5.49	7.016841 ±0.045008	98.24%	0.082114 ±0.000402	87.037 ±0.086
Ministral-3-14B-Reasoning-2512-Q3_K	3.5000	5.92	6.887992 ±0.043694	98.66%	0.061232 ±0.000311	88.897 ±0.081
Ministral-3-14B-Reasoning-2512-Q4_K_S	4.2500	7.18	6.527023 ±0.040267	99.66%	0.015459 ±0.000089	93.995 ±0.061
Ministral-3-14B-Reasoning-2512-Q4_K	4.5000	7.61	6.504096 ±0.040048	99.76%	0.011151 ±0.000065	94.870 ±0.057
Ministral-3-14B-Reasoning-2512-Q4_K_M-bartowski	4.8755	8.24	6.495948 ±0.039938	99.79%	0.009252 ±0.000054	95.480 ±0.053
Ministral-3-14B-Reasoning-2512-Q4_K_M-unsloth	4.8755	8.24	6.496198 ±0.039898	99.79%	0.009316 ±0.000056	95.457 ±0.053
Ministral-3-14B-Reasoning-2512-Q4_K_M-bpw	4.8755	8.24	6.479312 ±0.039858	99.83%	0.007521 ±0.000050	95.795 ±0.052
Ministral-3-14B-Reasoning-2512-Q5_K_S	5.2500	8.87	6.468715 ±0.039772	99.88%	0.005323 ±0.000034	96.391 ±0.048
Ministral-3-14B-Reasoning-2512-Q5_K	5.5000	9.29	6.457297 ±0.039602	99.92%	0.003445 ±0.000020	97.079 ±0.043
Ministral-3-14B-Reasoning-2512-Q6_K	6.4998	11.0	6.448191 ±0.039554	99.97%	0.001177 ±0.000007	98.333 ±0.033
Ministral-3-14B-Reasoning-2512-Q8_0	8.4999	14.4	6.439012 ±0.039469	99.99%	0.000137 ±0.000001	99.442 ±0.019

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 appropriate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Ministral-3-14B-Reasoning-2512-IQ1_L	46.6667	57.3333	31.8667	28.6667	56.4000	44.19
Ministral-3-14B-Reasoning-2512-IQ2_S	62.8000	69.3333	39.7333	28.8000	67.0667	53.55
Ministral-3-14B-Reasoning-2512-IQ2_XS	62.1333	68.6667	36.9333	28.4000	67.4667	52.72
Ministral-3-14B-Reasoning-2512-IQ2_XXS	57.3333	63.6000	34.8000	27.7333	64.0000	49.49
Ministral-3-14B-Reasoning-2512-IQ3_XXS	69.0667	75.3333	40.4000	32.4000	71.0667	57.65
Ministral-3-14B-Reasoning-2512-Q2_K	69.4667	72.2667	40.9333	28.5333	69.3333	56.11
Ministral-3-14B-Reasoning-2512-Q3_K_L	69.0667	77.2000	41.8667	32.5333	74.9333	59.12
Ministral-3-14B-Reasoning-2512-Q3_K_S	70.0000	75.8666	41.6000	32.8000	70.9333	58.24
Ministral-3-14B-Reasoning-2512-Q3_K	71.3333	76.0000	41.4667	31.6000	75.0667	59.09
Ministral-3-14B-Reasoning-2512-Q4_K_S	70.6667	78.0000	42.6667	32.8000	74.6667	59.76
Ministral-3-14B-Reasoning-2512-Q4_K	69.7333	77.8667	43.4667	32.2667	74.5333	59.57
Ministral-3-14B-Reasoning-2512-Q4_K_M-bartowski	69.2000	78.1333	43.6000	32.2667	74.8000	59.60
Ministral-3-14B-Reasoning-2512-Q4_K_M-unsloth	68.6667	78.6666	42.6667	30.5333	73.3333	58.77
Ministral-3-14B-Reasoning-2512-Q4_K_M-bpw	70.0000	78.2667	43.2000	32.4000	74.1333	59.60
Ministral-3-14B-Reasoning-2512-Q5_K_S	70.2667	78.4000	42.6667	32.0000	73.8667	59.44
Ministral-3-14B-Reasoning-2512-Q5_K	70.0000	78.8000	42.5333	31.2000	73.4667	59.20
Ministral-3-14B-Reasoning-2512-Q6_K	69.6000	79.0666	43.0667	31.6000	74.5333	59.57
Ministral-3-14B-Reasoning-2512-Q8_0	69.8667	78.6666	42.6667	31.6000	74.4000	59.44

Tokens per second benchmarks

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

model	size	params	backend	threads	test	t/s
Ministral-3-14B-Reasoning-2512-Q4_K_M-bpw	7.67 GiB	13.51 B	Metal,BLAS	12	pp512	507.26 ±0.86
Ministral-3-14B-Reasoning-2512-Q4_K_M-bpw	7.67 GiB	13.51 B	Metal,BLAS	12	tg128	41.11 ±1.35
Ministral-3-14B-Reasoning-2512-Q4_K_M-bpw	7.67 GiB	13.51 B	Metal,BLAS	12	pp1024+tg1024	68.68 ±0.18
Ministral-3-14B-Reasoning-2512-Q4_K_M-bartowski	7.67 GiB	13.51 B	Metal,BLAS	12	pp512	486.21 ±6.62
Ministral-3-14B-Reasoning-2512-Q4_K_M-bartowski	7.67 GiB	13.51 B	Metal,BLAS	12	tg128	48.05 ±3.19
Ministral-3-14B-Reasoning-2512-Q4_K_M-bartowski	7.67 GiB	13.51 B	Metal,BLAS	12	pp1024+tg1024	78.17 ±0.24
Ministral-3-14B-Reasoning-2512-Q4_K_M-unsloth	7.67 GiB	13.51 B	Metal,BLAS	12	pp512	501.23 ±12.03
Ministral-3-14B-Reasoning-2512-Q4_K_M-unsloth	7.67 GiB	13.51 B	Metal,BLAS	12	tg128	45.61 ±3.09
Ministral-3-14B-Reasoning-2512-Q4_K_M-unsloth	7.67 GiB	13.51 B	Metal,BLAS	12	pp1024+tg1024	78.73 ±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!

While I'm grateful to all contributors, I want to recognise three in particular:

• Colin Kealty (Bartowski), for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face
• Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries
• Iwan Kawrakow for being one of the key authors behind the many quantization algorithms and the imatrix functionality.