Mistral Large 4: official announcement source notes
Source: https://mistral.ai/news/mistral-large-4/
Publisher: Mistral
Published: October 6, 2026
Retrieved: October 8, 2026

Provenance: These are transcribed factual fields and paraphrased notes from the official announcement, read through an external web reader. This file is not a raw HTML snapshot or a hardware measurement. The Open-Science session did not retrieve the page itself; its access request was declined because that dialog offered only a persistent, cross-project domain grant.

Published facts relevant to memory estimation:
- Model: Mistral Large 4, also called ML4 or Le Chonk.
- Total parameters: 1 trillion.
- Active parameters: 52 billion.
- The model is natively multimodal.
- A public API preview was announced on October 6.
- The announcement says weights will be released by the end of October 2026.
- Further architecture information is planned alongside the weights. The announcement does not provide the exact layer count, KV-head count, head dimension, quantization layout, or deployment memory measurements needed to calculate exact runtime memory.

Analysis boundary: Treat 1 trillion as the published rounded parameter count, not a count from a downloaded checkpoint. Any KV-cache architecture, group size, scale/zero-point format, runtime reserve, offloading arrangement, or precision support introduced in calculations must be clearly labeled as an illustrative assumption, not an official Mistral Large 4 specification. No weights were downloaded and no model inference was run to produce these notes.
