On March 6, 2026, Mistral announced Mistral 3, its new flagship open multimodal family. The release covers dense 14B, 8B, and 3B models under an Apache 2.0 license, plus a new Mistral Large 3 model with 41B active parameters and 675B total parameters for teams that want a larger frontier-style option inside the same stack.

Mistral says the models are multilingual and multimodal, and that the family was trained from scratch rather than adapted from an earlier checkpoint. The company positions the release as a practical open alternative for organizations that want strong small-to-mid-size models without moving immediately to closed, API-only deployments. According to the announcement, Mistral 3 currently ranks as the #2 open-source non-reasoning model on LMArena, while a reasoning version is planned next.

Release highlights

Mistral 3 14B, 8B, and 3B are open under Apache 2.0.
The company says optimized checkpoints are available for Blackwell NVL72 and for a single 8xA100 or 8xH100 node through vLLM.
Mistral Large 3 is described as a larger model with 41B active and 675B total parameters.
Mistral says the launch builds toward a full family with Mistral Medium 3.1, Mistral Small 3.2, and Ministral 3.

The operational angle is important. Many teams do not need the biggest model possible; they need a model they can license clearly, deploy on hardware they already control, and tune for multimodal workloads without a large engineering burden. By shipping dense sizes from 3B to 14B and pairing them with a larger companion model, Mistral is covering both embedded deployments and more capable server-side workloads.

For the open-model market, the announcement keeps pressure on larger proprietary vendors. Mistral is not just offering another benchmark result; it is packaging licensing, hardware guidance, and a full product-family strategy around open multimodal inference.

#apache-2.0

Mistral launches Mistral 3 open multimodal family under Apache 2.0