jeffcookio/

Mistral-Small-3.1-24B-Instruct-2503-HF-gptqmodel-4b-128g

This is a Mistral-Small-3.1-24B-Instruct-2503 quantized from a hacked-up GPTQModel that has preliminary Mistral3ForConditionalGeneration support. There were several weird changes. Calibration was run against the flickr30k dataset (with too few samples; may upload a version with more significant calibration soon), and thus this should be a true vision-aware quant of the Mistral Small 3.1 HF checkpoint.

You need this branch of vLLM to run: https://github.com/sjuxax/vllm/tree/Mistral3.1

Another "feature" of this version is that it was quantized with a preliminary implementation of block-diagonal Hessians (which was authored entirely by Grok3). This allowed me to compute the quantization without OOM in my 24G VRAM.