Production model

Mistral Large 3 (675B)

Mistral AI · Released 2025-12-02 · 675B parameters (41B active) · 256,000 ctx · open weights

Architecture

Positional encoding	—
Norm placement	—
Norm type	—
QK-Norm	—
Activation	—
Attention	—
MoE	Granular MoE (exact expert count and routing not disclosed publicly)
Other	675B total, 41B active per token (39B language + 2.5B vision encoder), 256K context window, Native multimodal (text + image), Apache 2.0 license, Architecture disclosure beyond MoE/params/context not in the public model card at release

Techniques used

Sparsely-Gated MoE — 'Granular Mixture-of-Experts' per Mistral's announcement; exact expert count and routing top-K not disclosed in the public materials at release.

Mistral Large 3 (December 2025) is Mistral AI’s largest open-weights release: 675B total parameters with 41B active per token, 256K context, native multimodal. The model is shipped under Apache 2.0 — the broadest permissive license at this scale outside the DeepSeek V3 line.

The architectural disclosure at release is unusually thin for an open-weights frontier model:

Parameter shape: 675B total / 41B active, with the public breakdown placing 2.5B of the total on a vision encoder and ~673B / ~39B on the language model component.
MoE: described in the announcement as “granular Mixture-of-Experts.” Exact expert count, routing top-K, and presence/absence of a shared expert are not stated.
Context: 256K tokens.

Per the citation policy, this entry leaves the rest of the architecture fields null rather than infer them. No technical report or paper has been published for Mistral Large 3 beyond the announcement post and the Hugging Face model card; Mistral has stated training used roughly 3,000 NVIDIA H200 GPUs.

For Mistral’s earlier disclosed architecture see Mistral 7B and the Mixtral 8x7B entry.

Sources

Models like this

Nemotron 3 Nano 30B-A3BNVIDIA33% overlap
Qwen3 30B-A3BAlibaba (Qwen Team)17% overlap
Qwen3 235B-A22BAlibaba (Qwen Team)14% overlap