Mistral 7B

Mistral 7B is a dense autoregressive text language model from Mistral AI, a French company. It is a Transformer, but it uses sliding-window attention so each token attends within a moving local window, and grouped-query attention to reduce the cost of serving generated text. Those design choices distinguish it from a generic 7B Transformer without turning it into a mixture-of-experts model. Mixtral is the sparse expert model; Mistral 7B is not. TT-Metal lists Mistral 7B on a single n150 Wormhole device with batch 32. That modest one-chip scale contrasts with Mixtral 8x7B, which the same matrix partitions eight ways on a Wormhole QuietBox.