eaddario/DeepSeek-R1-Distill-Llama-8B-GGUF

Experimental GGUF quantized versions of deepseek-ai/DeepSeek-R1-Distill-Llama-8B

Using LLaMA C++ release b4930 for quantization.

Original model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B

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 that I'm using to produce these experimental versions is explained in Squeezing Tensor Bits: the quest for smaller LLMs, but at a high level it involves using a custom version of the llama-quantize tool to selectively quantize different tensors at different levels.

There’re two pull requests (#12511 & #12512) to merge these changes back into the core llama.cpp project. This may or may not ever happen but until then, the modified version will be available on my GitHub.

In addition to llama-quantize, there’s a version of llama-perplexity that allows you to continue generating test scores even if there’s a context window overflow (original behaviour is to stop).

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. Select an appropiate quant level for each tensor using a modified version of llama-quantize
5. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
6. Keep versions with the best scores
7. 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-Llama-8B-IQ3_M	3.78	N/A	3.48	7.9%
DeepSeek-R1-Distill-Llama-8B-IQ3_S	N/A	N/A	3.24	N/A
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	4.68	N/A	4.30	8.1%
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	4.32	N/A	3.45	20.1%
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	4.02	4.02	3.37	16.2%
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	3.66	N/A	3.28	10.4%
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.92	4.92	4.44	9.8%
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	4.69	N/A	4.31	8.1%
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	5.73	5.73	5.35	6.6%
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	5.60	N/A	5.19	7.3%
DeepSeek-R1-Distill-Llama-8B-Q6_K	6.60	6.60	6.17	6.5%
DeepSeek-R1-Distill-Llama-8B-Q8_0	8.54	8.54	7.84	8.2%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
DeepSeek-R1-Distill-Llama-8B-IQ3_M	18.513609 ±0.156607	91.75%	0.532214 ±0.001850	19.740 ±0.070
DeepSeek-R1-Distill-Llama-8B-IQ3_S	19.112490 ±0.165621	91.50%	0.547460 ±0.001918	19.434 ±0.071
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	16.368324 ±0.146856	96.44%	0.243864 ±0.001293	12.434 ±0.063
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	17.319338 ±0.145208	92.60%	0.477681 ±0.001783	18.029 ±0.069
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	17.594278 ±0.144215	91.45%	0.553217 ±0.001919	19.812 ±0.070
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	18.458827 ±0.153192	90.73%	0.602604 ±0.002075	20.369 ±0.072
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	14.675582 ±0.124114	98.02%	0.126483 ±0.000727	9.254 ±0.052
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	14.635869 ±0.123593	97.97%	0.130017 ±0.000732	9.361 ±0.053
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	14.310641 ±0.120564	98.83%	0.076779 ±0.000579	7.195 ±0.050
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	14.305297 ±0.120363	98.82%	0.077645 ±0.000576	7.256 ±0.049
DeepSeek-R1-Distill-Llama-8B-Q6_K	14.223713 ±0.119989	98.91%	0.071936 ±0.000599	6.892 ±0.051
DeepSeek-R1-Distill-Llama-8B-Q8_0	14.249305 ±0.120101	98.99%	0.066705 ±0.000592	6.662 ±0.052
DeepSeek-R1-Distill-Llama-8B-F16	14.009216 ±0.118474	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-Llama-8B-IQ3_M	51.3369 ±1.8288	69.33	32.9333 ±1.7172	32.2981 ±2.6100	64.2667 ±1.7510	50.03
DeepSeek-R1-Distill-Llama-8B-IQ3_S	48.1928 ±1.8294	66.67	34.5333 ±1.7374	31.1321 ±2.6007	65.2000 ±1.7405	49.15
DeepSeek-R1-Distill-Llama-8B-IQ4_NL	51.8072 ±1.8294	72.00	36.6667 ±1.7608	32.1212 ±2.5743	65.3333 ±1.7389	51.59
DeepSeek-R1-Distill-Llama-8B-Q3_K_L	51.4056 ±1.8299	65.20	34.5333 ±1.7374	33.8509 ±2.6412	68.0000 ±1.7045	50.60
DeepSeek-R1-Distill-Llama-8B-Q3_K_M	51.5395 ±1.8298	70.13	34.4000 ±1.7358	31.3665 ±2.5897	66.1333 ±1.7292	50.71
DeepSeek-R1-Distill-Llama-8B-Q3_K_S	50.8021 ±1.8292	71.20	31.3333 ±1.6949	34.5679 ±2.6462	67.4667 ±1.7119	51.07
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	52.8782 ±1.8276	74.40	35.7333 ±1.7510	34.0625 ±2.6534	67.4667 ±1.7119	52.91
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	50.3347 ±1.8306	74.40	34.8000 ±1.7405	37.1069 ±2.7133	69.4667 ±1.6828	53.22
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	52.4766 ±1.8284	73.20	33.2000 ±1.7207	36.0000 ±2.6667	68.4000 ±1.6988	52.66
DeepSeek-R1-Distill-Llama-8B-Q4_K_S	50.2008 ±1.8306	74.93	34.1333 ±1.7325	34.6875 ±2.6650	66.9333 ±1.7190	52.18
DeepSeek-R1-Distill-Llama-8B-Q5_K_M	53.6096 ±1.8246	72.66	34.6667 ±1.7389	35.7798 ±2.6549	66.0000 ±1.7309	52.54
DeepSeek-R1-Distill-Llama-8B-Q5_K_S	53.2798 ±1.8267	75.33	35.2000 ±1.7451	36.3354 ±2.6845	68.8000 ±1.6929	53.79
DeepSeek-R1-Distill-Llama-8B-Q6_K	51.8717 ±1.8281	76.00	33.3333 ±1.7225	35.1852 ±2.6571	68.6667 ±1.6949	53.01
DeepSeek-R1-Distill-Llama-8B-Q8_0	51.9359 ±1.8268	72.00	33.4667 ±1.7242	35.0000 ±2.6705	68.6667 ±1.6949	52.21
DeepSeek-R1-Distill-Llama-8B-F16	51.4706 ±1.8286	72.93	36.1333 ±1.7553	36.1111 ±2.6726	68.8000 ±1.6929	53.09

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-Llama-8B-Q4_K_M	4.13 GiB	8.03 B	Metal,BLAS	6	pp512	330.94 ± 1.42
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.13 GiB	8.03 B	Metal,BLAS	6	tg128	26.28 ± 0.11
DeepSeek-R1-Distill-Llama-8B-Q4_K_M	4.13 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.90 ± 0.05
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	329.03 ± 0.11
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	25.79 ± 0.92
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-bartowski	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.35 ± 0.93
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	328.93 ± 0.15
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	26.43 ± 0.01
DeepSeek-R1-Distill-Llama-8B-Q4_K_M-unsloth	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.38 ± 0.97

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.