Mistral AI released a 7.3 billion parameter model in September 2023 that outperforms Llama 2 13B across most benchmarks and approaches Llama 34B on reasoning, STEM, and comprehension tasks. The model uses Grouped Query Attention and Sliding Window Attention to deliver approximately 2x inference speed gain on 16k token sequences while reducing cache memory by half, and was released under Apache 2.0 license for unrestricted use. For financial institutions, this establishes an empirical baseline showing that architectural efficiency can substitute materially for model scale, with implications for inference-cost structure and competitive positioning among open-weight models.
223 other tracked developments touching the same claims.