Llamacpp imatrix Quantizations of Devstral-2-123B-Instruct-2512 by mistralai
Using llama.cpp release b7335 for quantization.
Original model: https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512
Thank you to ngxson and compilade for helping to get the conversion from FP8 working correctly
Do not expect this to work with mistral-vibe
More work is likely needed to get llama.cpp server to work correctly with mistral-vibe tool calling, in quick testing tool calls work but only when they're singular, chained tool calls break
ANY issues should be reported to llama.cpp, NOT mistral!
All quants made using imatrix option with dataset from here combined with a subset of combined_all_small.parquet from Ed Addario here
Run them in LM Studio
Run them directly with llama.cpp, or any other llama.cpp based project
Prompt format
<s>[SYSTEM_PROMPT]{system_prompt}[/SYSTEM_PROMPT][INST]{prompt}[/INST]
Download a file (not the whole branch) from below:
| Filename | Quant type | File Size | Split | Description |
|---|---|---|---|---|
| Devstral-2-123B-Instruct-2512-Q8_0.gguf | Q8_0 | 132.85GB | true | Extremely high quality, generally unneeded but max available quant. |
| Devstral-2-123B-Instruct-2512-Q6_K.gguf | Q6_K | 102.58GB | true | Very high quality, near perfect, recommended. |
| Devstral-2-123B-Instruct-2512-Q5_K_M.gguf | Q5_K_M | 88.32GB | true | High quality, recommended. |
| Devstral-2-123B-Instruct-2512-Q5_K_S.gguf | Q5_K_S | 86.18GB | true | High quality, recommended. |
| Devstral-2-123B-Instruct-2512-Q4_1.gguf | Q4_1 | 78.47GB | true | Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon. |
| Devstral-2-123B-Instruct-2512-Q4_K_L.gguf | Q4_K_L | 76.09GB | true | Uses Q8_0 for embed and output weights. Good quality, recommended. |
| Devstral-2-123B-Instruct-2512-Q4_K_M.gguf | Q4_K_M | 74.90GB | true | Good quality, default size for most use cases, recommended. |
| Devstral-2-123B-Instruct-2512-Q4_K_S.gguf | Q4_K_S | 71.25GB | true | Slightly lower quality with more space savings, recommended. |
| Devstral-2-123B-Instruct-2512-Q4_0.gguf | Q4_0 | 71.00GB | true | Legacy format, offers online repacking for ARM and AVX CPU inference. |
| Devstral-2-123B-Instruct-2512-IQ4_NL.gguf | IQ4_NL | 70.90GB | true | Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference. |
| Devstral-2-123B-Instruct-2512-Q3_K_XL.gguf | Q3_K_XL | 67.48GB | true | Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability. |
| Devstral-2-123B-Instruct-2512-IQ4_XS.gguf | IQ4_XS | 67.07GB | true | Decent quality, smaller than Q4_K_S with similar performance, recommended. |
| Devstral-2-123B-Instruct-2512-Q3_K_L.gguf | Q3_K_L | 66.07GB | true | Lower quality but usable, good for low RAM availability. |
| Devstral-2-123B-Instruct-2512-Q3_K_M.gguf | Q3_K_M | 60.62GB | true | Low quality. |
| Devstral-2-123B-Instruct-2512-IQ3_M.gguf | IQ3_M | 56.79GB | true | Medium-low quality, new method with decent performance comparable to Q3_K_M. |
| Devstral-2-123B-Instruct-2512-Q3_K_S.gguf | Q3_K_S | 54.37GB | true | Low quality, not recommended. |
| Devstral-2-123B-Instruct-2512-IQ3_XS.gguf | IQ3_XS | 51.66GB | true | Lower quality, new method with decent performance, slightly better than Q3_K_S. |
| Devstral-2-123B-Instruct-2512-IQ3_XXS.gguf | IQ3_XXS | 48.37GB | false | Lower quality, new method with decent performance, comparable to Q3 quants. |
| Devstral-2-123B-Instruct-2512-Q2_K_L.gguf | Q2_K_L | 48.16GB | false | Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable. |
| Devstral-2-123B-Instruct-2512-Q2_K.gguf | Q2_K | 46.59GB | false | Very low quality but surprisingly usable. |
| Devstral-2-123B-Instruct-2512-IQ2_M.gguf | IQ2_M | 42.98GB | false | Relatively low quality, uses SOTA techniques to be surprisingly usable. |
| Devstral-2-123B-Instruct-2512-IQ2_S.gguf | IQ2_S | 39.74GB | false | Low quality, uses SOTA techniques to be usable. |
| Devstral-2-123B-Instruct-2512-IQ2_XS.gguf | IQ2_XS | 37.32GB | false | Low quality, uses SOTA techniques to be usable. |
| Devstral-2-123B-Instruct-2512-IQ2_XXS.gguf | IQ2_XXS | 33.67GB | false | Very low quality, uses SOTA techniques to be usable. |
| Devstral-2-123B-Instruct-2512-IQ1_M.gguf | IQ1_M | 29.62GB | false | Extremely low quality, not recommended. |
| Devstral-2-123B-Instruct-2512-IQ1_S.gguf | IQ1_S | 27.19GB | false | 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
- 842
Model tree for bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF
Base model
mistralai/Devstral-2-123B-Instruct-2512