neuralmagic/Mistral-7B-Instruct-v0.3-GPTQ-4bit

Model Card for Mistral-7B-Instruct-v0.3 quantized to 4bit weights

• Weight-only quantization of Mistral-7B-Instruct-v0.3 via GPTQ to 4bits with group_size=128
• GPTQ optimized for 99.75% accuracy recovery relative to the unquantized model

Open LLM Leaderboard evaluation scores

	Mistral-7B-Instruct-v0.3	Mistral-7B-Instruct-v0.3-GPTQ-4bit (this model)
arc-c 25-shot	63.48	63.40
mmlu 5-shot	61.13	60.89
hellaswag 10-shot	84.49	84.04
winogrande 5-shot	79.16	79.08
gsm8k 5-shot	43.37	45.41
truthfulqa 0-shot	59.65	57.48
Average Accuracy	65.21	65.05
Recovery	100%	99.75%

vLLM Inference Performance

This model is ready for optimized inference using the Marlin mixed-precision kernels in vLLM: https://github.com/vllm-project/vllm

Simply start this model as an inference server with:

python -m vllm.entrypoints.openai.api_server --model neuralmagic/Mistral-7B-Instruct-v0.3-GPTQ-4bit