Papers · Mistral 7B
AI PAPER

Mistral 7B

Released a 7B parameter model that outperformed LLaMA 2 13B on all benchmarks through architectural innovations including grouped-query attention and sliding window attention.

Authors
Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch +7
Institution
Mistral AI
Published
October NaN, 2023
Citations
2,500
Category
Architecture
Impact
notable
PAPER EXPLAINED

Plain-English Summary

Mistral AI’s debut model punched far above its weight class. At just 7 billion parameters, it outperformed Meta’s LLaMA 2 at 13B parameters on every benchmark tested. It even approached LLaMA 2 34B on several tasks. The secret was a combination of architectural innovations — sliding window attention for efficient long-context processing and grouped-query attention for faster inference — plus aggressive training on high-quality data.

The model was released under the Apache 2.0 license with no usage restrictions, the most permissive release of a competitive language model at the time. Mistral dropped the model weights via a BitTorrent magnet link with no announcement, establishing the startup’s distinctive approach to releases.

Key Innovation

Sliding window attention limits each token to attending only to the previous 4,096 tokens within each attention layer, while allowing information to propagate further through multiple layers. This provides effectively unlimited context propagation while keeping per-layer attention cost linear rather than quadratic.

Grouped-query attention (GQA) shares key-value heads across multiple query heads, reducing the memory footprint of the KV-cache and enabling faster inference with minimal quality loss.

Impact on the Field

Mistral 7B demonstrated that a well-funded startup with experienced researchers could compete with trillion-dollar companies on model quality. The unconditional open-source release (no registration, no usage restrictions) pressured other labs to be more permissive with their licensing.

The model became the most popular base for fine-tuning in the open-source community, eventually competing with LLaMA for the default foundation model role.

Models That Built on This

Mixtral 8x7B extended the architecture with mixture-of-experts. Zephyr used DPO on Mistral 7B for efficient chat training. Thousands of community fine-tunes were built on Mistral 7B. Mistral AI subsequently released Medium and Large models, building a full model family from this foundation. The sliding window attention technique has been adopted by subsequent models from other organizations.