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

Click to view download instructions

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF --include "mistralai_Devstral-2-123B-Instruct-2512-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF --include "mistralai_Devstral-2-123B-Instruct-2512-Q8_0/*" --local-dir ./

You can either specify a new local-dir (mistralai_Devstral-2-123B-Instruct-2512-Q8_0) or download them all in place (./)

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.

Click to view Q4_0_X_X information (deprecated

I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.

Click to view benchmarks on an AVX2 system (EPYC7702)

model	size	params	backend	threads	test	t/s	% (vs Q4_0)
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp512	204.03 ± 1.03	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp1024	282.92 ± 0.19	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp2048	259.49 ± 0.44	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg128	39.12 ± 0.27	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg256	39.31 ± 0.69	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg512	40.52 ± 0.03	100%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp512	301.02 ± 1.74	147%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp1024	287.23 ± 0.20	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp2048	262.77 ± 1.81	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg128	18.80 ± 0.99	48%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg256	24.46 ± 3.04	83%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg512	36.32 ± 3.59	90%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp512	271.71 ± 3.53	133%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp1024	279.86 ± 45.63	100%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp2048	320.77 ± 5.00	124%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg128	43.51 ± 0.05	111%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg256	43.35 ± 0.09	110%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg512	42.60 ± 0.31	105%

Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation

Which file should I choose?

Click here for details

A great write up with charts showing various performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.

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