eaddario/gemma-3-12b-it-GGUF

Experimental layer-wise quantization of google/gemma-3-12b-it

Using LLaMA C++ release b5490 for quantization.

Original model: google/gemma-3-12b-it

From the original model creators:

Terms of Use: Terms

Authors: Google DeepMind

Model Information Summary description and brief definition of inputs and outputs.

Description Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants. Gemma 3 has a large, 128K context window, multilingual support in over 140 languages, and is available in more sizes than previous versions. Gemma 3 models are well-suited for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Inputs and outputs Input:

Text string, such as a question, a prompt, or a document to be summarized Images, normalized to 896 x 896 resolution and encoded to 256 tokens each Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B size Output:

Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document Total output context of 8192 tokens Usage

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 and llama-quantize to identify influential tensors, and quantize the most important layers to higher bit precision and the less important to lower bits. 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.

As of version b5125 llama-quantize can now 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).

The modified version of llama-imatrix generates useful statistics to guide the tensor selection process, --show-statistics will display:

• Σ(Bias): the sum of all activations over the tensor (i.e. the Importance Scores)
• Min & Max: minimum and maximum activation values
• μ & σ: activations' mean and standard deviation
• % Active: proportion of elements whose average activation exceeds a very small threshold (1e-6). Helpful to determine how alive/dormant the tensor is during inference
• N: number of activations in the tensor
• Entropy: entropy of the 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 individial 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(Bias) 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 when 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. In this case however, whilst both have versions of this model, Unsloth's uses a different vocabulary size on their quants (262144 vs 262208) which makes a like-for-like comparison invalid.

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 modeled, 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 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 modified version of llama-imatrix
5. Select an appropiate quant level for each tensor and quantize the model using llama-quantize
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	Repo	Shrinkage
gemma-3-12b-it-IQ3_M	5.66	5.41	4.4%
gemma-3-12b-it-IQ3_S	5.21	5.23	-0.4%
gemma-3-12b-it-IQ4_NL	6.89	6.39	7.3%
gemma-3-12b-it-Q3_K_L	6.48	5.52	14.8%
gemma-3-12b-it-Q3_K_M	6.01	5.22	13.1%
gemma-3-12b-it-Q3_K_S	5.46	4.99	8.6%
gemma-3-12b-it-Q4_K_M	7.30	6.43	11.9%
gemma-3-12b-it-Q4_K_S	6.94	6.40	7.8%
gemma-3-12b-it-Q5_K_M	8.44	7.61	9.8%
gemma-3-12b-it-Q5_K_S	8.23	7.58	7.9%
gemma-3-12b-it-Q6_K	9.66	9.37	3.0%
gemma-3-12b-it-Q8_0	12.50	11.40	8.8%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
gemma-3-12b-it-IQ3_M	9.819911 ±0.080440	95.67%	0.189526 ±0.001269	12.497 ±0.068
gemma-3-12b-it-IQ3_S	9.844402 ±0.079485	94.86%	0.232563 ±0.001411	13.899 ±0.069
gemma-3-12b-it-IQ4_NL	9.784609 ±0.080899	97.10%	0.121353 ±0.000786	10.126 ±0.059
gemma-3-12b-it-Q3_K_L	10.193888 ±0.083697	94.26%	0.261491 ±0.001586	14.559 ±0.072
gemma-3-12b-it-Q3_K_M	9.860796 ±0.078875	93.95%	0.273903 ±0.001588	15.007 ±0.072
gemma-3-12b-it-Q3_K_S	10.351147 ±0.082692	92.32%	0.360147 ±0.001863	17.074 ±0.072
gemma-3-12b-it-Q4_K_M	9.788871 ±0.080127	97.05%	0.127527 ±0.000807	10.439 ±0.060
gemma-3-12b-it-Q4_K_M-bartowski	9.197963 ±0.073909	98.63%	0.045560 ±0.000416	6.400 ±0.054
gemma-3-12b-it-Q4_K_S	9.791993 ±0.080114	97.02%	0.129268 ±0.000813	10.528 ±0.060
gemma-3-12b-it-Q5_K_M	9.492645 ±0.076969	98.44%	0.060229 ±0.000376	7.340 ±0.046
gemma-3-12b-it-Q5_K_S	9.495540 ±0.077022	98.44%	0.060458 ±0.000376	7.343 ±0.045
gemma-3-12b-it-Q6_K	9.380805 ±0.075743	98.88%	0.041053 ±0.000229	6.061 ±0.037
gemma-3-12b-it-Q8_0	9.350086 ±0.075331	99.00%	0.036289 ±0.000190	5.712 ±0.034
gemma-3-12b-it-F16	9.042786 ±0.071503	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
gemma-3-12b-it-IQ3_M	69.0667 +/- 1.6889	79.60	43.0667 +/- 1.8093	36.9333 +/- 1.7635	73.7333 +/- 1.6080	60.48
gemma-3-12b-it-IQ3_S	67.2000 +/- 1.7155	79.20	43.4667 +/- 1.8113	40.4000 +/- 1.7930	73.2000 +/- 1.6184	60.69
gemma-3-12b-it-IQ4_NL	68.8000 +/- 1.6929	79.06	43.0667 +/- 1.8093	39.0667 +/- 1.7827	73.7333 +/- 1.6080	60.75
gemma-3-12b-it-Q3_K_L	68.6667 +/- 1.6949	78.26	43.4667 +/- 1.8113	39.4667 +/- 1.7860	72.6667 +/- 1.6284	60.51
gemma-3-12b-it-Q3_K_M	68.0000 +/- 1.7045	78.93	42.9333 +/- 1.8086	39.8667 +/- 1.7890	71.7333 +/- 1.6453	60.29
gemma-3-12b-it-Q3_K_S	66.9333 +/- 1.7190	78.80	42.9333 +/- 1.8086	40.0000 +/- 1.7900	70.2667 +/- 1.6702	59.79
gemma-3-12b-it-Q4_K_M	70.6667 +/- 1.6636	80.67	42.8000 +/- 1.8079	40.1333 +/- 1.7910	74.0000 +/- 1.6027	61.65
gemma-3-12b-it-Q4_K_M-bartowski	69.6000 +/- 1.6807	81.07	43.6000 +/- 1.8119	41.0667 +/- 1.7976	75.7333 +/- 1.5664	62.21
gemma-3-12b-it-Q4_K_S	70.2667 +/- 1.6702	80.67	43.0667 +/- 1.8093	40.2667 +/- 1.7920	74.4000 +/- 1.5947	61.73
gemma-3-12b-it-Q5_K_M	68.4000 +/- 1.6988	81.47	45.0667 +/- 1.8180	40.2667 +/- 1.7920	75.3333 +/- 1.5751	62.11
gemma-3-12b-it-Q5_K_S	68.6667 +/- 1.6949	81.46	44.5333 +/- 1.8160	40.1333 +/- 1.7910	74.9333 +/- 1.5836	61.95
gemma-3-12b-it-Q6_K	68.9333 +/- 1.6909	81.07	44.4000 +/- 1.8155	41.0667 +/- 1.7976	75.0667 +/- 1.5808	62.11
gemma-3-12b-it-Q8_0	68.0000 +/- 1.7045	80.93	44.0000 +/- 1.8138	40.9333 +/- 1.7967	75.4667 +/- 1.5722	61.87
gemma-3-12b-it-F16	69.2000 +/- 1.6869	81.20	45.3333 +/- 1.8190	41.4667 +/- 1.8002	74.8000 +/- 1.5864	62.40

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
gemma-3-12b-it-Q4_K_M	5.98 GiB	11.77 B	Metal,BLAS	12	pp512	504.86 ± 1.32
gemma-3-12b-it-Q4_K_M	5.98 GiB	11.77 B	Metal,BLAS	12	tg128	44.13 ± 0.18
gemma-3-12b-it-Q4_K_M	5.98 GiB	11.77 B	Metal,BLAS	12	pp1024+tg1024	72.53 ± 0.37
gemma-3-12b-it-Q4_K_M-bartowski	6.79 GiB	11.77 B	Metal,BLAS	12	pp512	530.98 ± 0.51
gemma-3-12b-it-Q4_K_M-bartowski	6.79 GiB	11.77 B	Metal,BLAS	12	tg128	45.54 ± 0.16
gemma-3-12b-it-Q4_K_M-bartowski	6.79 GiB	11.77 B	Metal,BLAS	12	pp1024+tg1024	74.36 ± 0.41

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

A big Thank You! to Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available on Huggingface, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries.