• Sources: primary, docs, discussion
  • Summary: Mistral released Mistral Large 4 in preview, a multimodal mixture-of-experts model that Mistral's documentation puts at 1.05T total parameters with 52B active, a 1.6B vision encoder and a 1M token context window, priced at $1.36 per million input tokens and $4.18 per million output tokens with a discounted tier at $0.68 and $2.09. Mistral states that the weights will be published at the end of October 2026, with an on-premise deployment path. The benchmark figures in the announcement are Mistral's own or relayed by Mistral from named third parties, and none were verified at source.
  • Why it matters: Sparse activation of 52B of the 1.05T parameters, plus the promised weights and on-premise path, would put a 1M-context model in range for teams that cannot send data to a hosted API, and the promise is dated rather than shipped.
  • Follow-up: Confirm whether the weights and their license ship by the end of October 2026.