TL
TL;DR
This work introduces Mistral 7B v0.1, a 7-billion-parameter language model engineered for superior performance and efficiency, which leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost.
Abstract
We introduce Mistral 7B v0.1, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7B outperforms Llama 2 13B across all evaluated benchmarks, and Llama 1 34B in reasoning, mathematics, and code generation. Our model leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. We also provide a model fine-tuned to follow instructions, Mistral 7B -- Instruct, that surpasses the Llama 2 13B -- Chat model both on human and automated benchmarks. Our models are released under the Apache 2.0 license.
References29 items
1
Efficient Memory Management for Large Language Model Serving with PagedAttention
2
Code Llama: Open Foundation Models for Code
3
Llama 2: Open Foundation and Fine-Tuned Chat Models
4
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
5