Llamacpp imatrix Quantizations of DeepSeek-V3-0324 by deepseek-ai
Using llama.cpp release b4944 for quantization.
Original model: https://huggingface.co/deepseek-ai/DeepSeek-V3-0324
All quants made using imatrix option with dataset from here
Run them in LM Studio
Run them directly with llama.cpp, or any other llama.cpp based project
Prompt format
<｜begin▁of▁sentence｜>{system_prompt}<｜User｜>{prompt}<｜Assistant｜><｜end▁of▁sentence｜><｜Assistant｜>
V2 uploads
I finally decided to go through llama-quant.cpp and update some of the tensor types, especially for MoE models, since they've kind of been left as-is since the original Mixtral.
These changes overall apply a bit more logic to the types, bumping a few values here and there across the board. These changes seem to have an overall positive impact on the results. They're similar to what Unsloth accomplished but they're in a more generic (and hopefully upstreamable) way.
The V2 weights uploaded are now final, I will delete the old ones if/when my PR is merged: https://github.com/ggml-org/llama.cpp/pull/12727
Download a file (not the whole branch) from below:
| Filename | Quant type | File Size | Split | Description |
|---|---|---|---|---|
| DeepSeek-V3-0324-Q8_0.gguf | Q8_0 | 713.29GB | true | Extremely high quality, generally unneeded but max available quant. |
| DeepSeek-V3-0324-Q6_K.gguf | Q6_K | 550.80GB | true | Very high quality, near perfect, recommended. |
| DeepSeek-V3-0324-Q5_K_M.gguf | Q5_K_M | 475.40GB | true | High quality, recommended. |
| DeepSeek-V3-0324-Q5_K_S.gguf | Q5_K_S | 461.81GB | true | High quality, recommended. |
| DeepSeek-V3-0324-Q4_1.gguf | Q4_1 | 419.94GB | true | Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon. |
| DeepSeek-V3-0324-Q4_K_M-V2.gguf | Q4_K_M | 406.99GB | true | Attempted to modify tensor quant levels for better performance. recommended |
| DeepSeek-V3-0324-Q4_K_M.gguf | Q4_K_M | 404.43GB | true | Good quality, default size for most use cases, recommended. |
| DeepSeek-V3-0324-Q4_K_S.gguf | Q4_K_S | 380.00GB | true | Slightly lower quality with more space savings, recommended. |
| DeepSeek-V3-0324-Q4_0.gguf | Q4_0 | 379.03GB | true | Legacy format, offers online repacking for ARM and AVX CPU inference. |
| DeepSeek-V3-0324-IQ4_NL.gguf | IQ4_NL | 378.07GB | true | Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference. |
| DeepSeek-V3-0324-IQ4_XS.gguf | IQ4_XS | 357.13GB | true | Decent quality, smaller than Q4_K_S with similar performance, recommended. |
| DeepSeek-V3-0324-Q3_K_XL.gguf | Q3_K_XL | 348.26GB | true | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
| DeepSeek-V3-0324-Q3_K_M.gguf | Q3_K_M | 319.20GB | true | Low quality. |
| DeepSeek-V3-0324-IQ3_XXS-V2.gguf | IQ3_XXS | 261.74GB | true | Attempted to modify tensor quant levels for better performance. |
| DeepSeek-V3-0324-IQ3_XXS.gguf | IQ3_XXS | 257.93GB | true | Lower quality, new method with decent performance, comparable to Q3 quants. |
| DeepSeek-V3-0324-Q2_K_L-V2.gguf | Q2_K_L | 247.35GB | true | Attempted to modify tensor quant levels for better performance. Also uses Q8_0 for embed and output weights. |
| DeepSeek-V3-0324-Q2_K_L.gguf | Q2_K_L | 244.93GB | true | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
| DeepSeek-V3-0324-IQ2_M-V2.gguf | IQ2_M | 224.49GB | true | Attempted to modify tensor quant levels for better performance. |
| DeepSeek-V3-0324-IQ2_M.gguf | IQ2_M | 217.43GB | true | Relatively low quality, uses SOTA techniques to be surprisingly usable. |
| DeepSeek-V3-0324-IQ2_S.gguf | IQ2_S | 197.00GB | true | Low quality, uses SOTA techniques to be usable. |
| DeepSeek-V3-0324-IQ2_XXS-V2.gguf | IQ2_XXS | 188.95GB | true | Attempted to modify tensor quant levels for better performance. |
| DeepSeek-V3-0324-IQ2_XXS.gguf | IQ2_XXS | 174.43GB | true | Very low quality, uses SOTA techniques to be usable. |
| DeepSeek-V3-0324-IQ1_M-V2.gguf | IQ1_M | 154.78GB | true | Attempted to modify tensor quant levels for better performance. Extremely low quality, not recommended. |
| DeepSeek-V3-0324-IQ1_M.gguf | IQ1_M | 148.88GB | true | Extremely low quality, not recommended. |
| DeepSeek-V3-0324-IQ1_S.gguf | IQ1_S | 133.56GB | true | Extremely low quality, not recommended. |
Embed/output weights
Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.
Downloading using huggingface-cli
ARM/AVX information
Previously, you would download Q4_0_4_4/4_8/8_8, and these would have their weights interleaved in memory in order to improve performance on ARM and AVX machines by loading up more data in one pass.
Now, however, there is something called "online repacking" for weights. details in this PR. If you use Q4_0 and your hardware would benefit from repacking weights, it will do it automatically on the fly.
As of llama.cpp build b4282 you will not be able to run the Q4_0_X_X files and will instead need to use Q4_0.
Additionally, if you want to get slightly better quality for , you can use IQ4_NL thanks to this PR which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed incrase.
Which file should I choose?
Credits
Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset.
Thank you ZeroWw for the inspiration to experiment with embed/output.
Thank you to LM Studio for sponsoring my work.
Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski
- Downloads last month
- 749
1-bit
2-bit
3-bit
4-bit
5-bit
6-bit
8-bit