eaddario/DeepSeek-R1-Distill-Qwen-7B-GGUF

Experimental layer-wise quantization of deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

Using LLaMA C++ release b5120 for quantization.

Original model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

From the original model creators:

DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. To support the research community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six dense models distilled from DeepSeek-R1 based on Llama and Qwen. DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models.

NOTE: Before running DeepSeek-R1 series models locally, we kindly recommend reviewing the Usage Recommendation section.

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 custom versions of llama-imatrix and llama-quantize to identify the 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.

There’re two pull requests (imatrix & quantize) 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 Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below).

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 a modified version of llama-imatrix
5. Select an appropiate quant level for each tensor using a modified version of 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	Unsloth	Repo	Shrinkage
DeepSeek-R1-Distill-Qwen-7B-IQ3_M	3.57	N/A	3.48	2.5%
DeepSeek-R1-Distill-Qwen-7B-IQ3_S	N/A	N/A	3.26	N/A
DeepSeek-R1-Distill-Qwen-7B-IQ4_NL	4.44	N/A	4.18	5.9%
DeepSeek-R1-Distill-Qwen-7B-Q3_K_L	4.09	N/A	3.54	13.4%
DeepSeek-R1-Distill-Qwen-7B-Q3_K_M	3.81	3.81	3.37	11.5%
DeepSeek-R1-Distill-Qwen-7B-Q3_K_S	3.49	N/A	3.14	10.0%
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	4.68	4.68	4.18	10.7%
DeepSeek-R1-Distill-Qwen-7B-Q4_K_S	4.46	N/A	4.06	9.0%
DeepSeek-R1-Distill-Qwen-7B-Q5_K_M	5.44	5.44	5.09	6.4%
DeepSeek-R1-Distill-Qwen-7B-Q5_K_S	5.32	N/A	4.97	6.6%
DeepSeek-R1-Distill-Qwen-7B-Q6_K	6.25	6.25	6.23	0.3%
DeepSeek-R1-Distill-Qwen-7B-Q8_0	8.10	8.10	7.27	10.2%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
DeepSeek-R1-Distill-Qwen-7B-IQ3_M	28.740047 ±0.291290	97.19%	0.229742 ±0.000770	11.793 ±0.050
DeepSeek-R1-Distill-Qwen-7B-IQ3_S	30.290800 ±0.307742	96.32%	0.310982 ±0.001014	13.415 ±0.057
DeepSeek-R1-Distill-Qwen-7B-IQ4_NL	23.570503 ±0.226124	98.59%	0.102854 ±0.000465	8.080 ±0.046
DeepSeek-R1-Distill-Qwen-7B-Q3_K_L	24.160705 ±0.229989	97.75%	0.173336 ±0.000603	10.337 ±0.048
DeepSeek-R1-Distill-Qwen-7B-Q3_K_M	24.967196 ±0.239198	97.50%	0.194299 ±0.000681	10.877 ±0.050
DeepSeek-R1-Distill-Qwen-7B-Q3_K_S	25.661098 ±0.246635	96.84%	0.243850 ±0.000852	12.143 ±0.054
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	23.125382 ±0.221860	99.24%	0.053997 ±0.000215	5.795 ±0.032
DeepSeek-R1-Distill-Qwen-7B-Q4_K_S	23.156199 ±0.222000	99.18%	0.058337 ±0.000233	6.026 ±0.034
DeepSeek-R1-Distill-Qwen-7B-Q5_K_M	22.726887 ±0.217691	99.75%	0.013562 ±0.000062	2.924 ±0.020
DeepSeek-R1-Distill-Qwen-7B-Q5_K_S	22.766826 ±0.218244	99.74%	0.014589 ±0.000070	3.024 ±0.020
DeepSeek-R1-Distill-Qwen-7B-Q6_K	22.859294 ±0.219461	99.87%	0.004317 ±0.000022	1.682 ±0.016
DeepSeek-R1-Distill-Qwen-7B-Q8_0	22.840693 ±0.219408	99.90%	0.001614 ±0.000011	1.050 ±0.010
DeepSeek-R1-Distill-Qwen-7B-F16	22.656280 ±0.216110	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
DeepSeek-R1-Distill-Qwen-7B-IQ3_M	50.4000 ±1.8269	55.73	31.3333 ±1.6949	26.4000 ±1.6106	59.7333 ±1.7920	45.78
DeepSeek-R1-Distill-Qwen-7B-IQ3_S	43.8667 ±1.8132	53.87	30.5333 ±1.6828	26.2667 ±1.6080	57.8667 ±1.8042	44.51
DeepSeek-R1-Distill-Qwen-7B-IQ4_NL	50.6667 ±1.8268	57.60	32.8000 ±1.7155	28.2667 ±1.6453	60.5333 ±1.7860	46.24
DeepSeek-R1-Distill-Qwen-7B-Q3_K_L	48.5333 ±1.8262	56.93	30.4000 ±1.6807	27.0667 ±1.6235	59.8667 ±1.7910	46.49
DeepSeek-R1-Distill-Qwen-7B-Q3_K_M	48.1333 ±1.8257	56.27	29.8667 ±1.6723	27.3333 ±1.6284	59.8667 ±1.7910	46.28
DeepSeek-R1-Distill-Qwen-7B-Q3_K_S	48.0000 ±1.8255	55.47	28.5333 ±1.6500	26.0000 ±1.6027	59.2000 ±1.7958	46.03
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	52.5333 ±1.8246	58.53	32.1333 ±1.7063	26.2667 ±1.6080	60.9333 ±1.7827	45.39
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-bartowski	50.9358 ±1.8291	60.67	35.8667 ±1.7525	28.7037 ±2.5171	61.0667 ±1.7816	47.45
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-unsloth	50.5348 ±1.8293	57.07	31.0667 ±1.6909	29.5031 ±2.5455	61.2000 ±1.7805	45.87
DeepSeek-R1-Distill-Qwen-7B-Q4_K_S	52.4000 ±1.8249	58.40	31.8667 ±1.7026	25.8667 ±1.6001	61.4667 ±1.7783	46.84
DeepSeek-R1-Distill-Qwen-7B-Q5_K_M	52.8000 ±1.8241	59.33	32.6667 ±1.7137	27.2000 ±1.6260	60.1333 ±1.7890	46.40
DeepSeek-R1-Distill-Qwen-7B-Q5_K_S	52.9333 ±1.8238	59.60	32.8000 ±1.7155	27.0667 ±1.6235	61.2000 ±1.7805	46.02
DeepSeek-R1-Distill-Qwen-7B-Q6_K	53.0667 ±1.8235	59.07	31.3333 ±1.6949	26.9333 ±1.6209	60.5333 ±1.7860	47.85
DeepSeek-R1-Distill-Qwen-7B-Q8_0	53.0667 ±1.8235	58.67	32.2667 ±1.7082	26.9333 ±1.6209	60.2667 ±1.7880	45.29
DeepSeek-R1-Distill-Qwen-7B-F16	52.9333 ±1.8238	58.67	32.1333 ±1.7063	26.6667 ±1.6158	60.0000 ±1.7900	46.81

Tokens per Second - Benchmarks

Scores generated using llama-bench. Q4_K_M quantizations from Bartowski and Unsloth included for comparison.

model	size	params	backend	threads	test	t/s
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	3.89 GiB	7.62 B	Metal,BLAS	6	pp512	333.33 ± 3.02
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	3.89 GiB	7.62 B	Metal,BLAS	6	tg128	29.14 ± 0.16
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M	3.89 GiB	7.62 B	Metal,BLAS	6	pp1024+tg1024	48.07 ± 0.09
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-bartowski	4.36 GiB	7.62 B	Metal,BLAS	6	pp512	354.27 ± 0.64
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-bartowski	4.36 GiB	7.62 B	Metal,BLAS	6	tg128	27.95 ± 0.04
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-bartowski	4.36 GiB	7.62 B	Metal,BLAS	6	pp1024+tg1024	46.60 ± 0.16
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-unsloth	4.36 GiB	7.62 B	Metal,BLAS	6	pp512	352.93 ± 0.92
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-unsloth	4.36 GiB	7.62 B	Metal,BLAS	6	tg128	27.66 ± 0.08
DeepSeek-R1-Distill-Qwen-7B-Q4_K_M-unsloth	4.36 GiB	7.62 B	Metal,BLAS	6	pp1024+tg1024	46.40 ± 0.45

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 in Hugginface, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries.