Source: Mistral Large 4 model card (Mistral AI, Hugging Face)
Model card for Mistral Large 4: a ~1T-parameter multimodal MoE with ~52B active, 1M context, under a modified-MIT licence restricting MaaS.
Contents
- What: Hugging Face model card
- Publisher: Mistral AI
- Date: 2026-09-30 (repository creation); accessed 2026-11-14
- Original: Mistral-Large-4-1T-A52B
Summary
Mistral’s largest model: trained from scratch in its European datacentres, released as open weights with an NVFP4 variant and an embedded speculative-decoding head.
Key claims
- About 1T total parameters, 49B active per token (52B including embeddings/output), 128 routed experts (4 active) plus 1 shared.
- Natively multimodal, 1M-token context, fluent in 160+ languages including all EU official languages.
- Ships an embedded DSpark draft head (3 dense MLA layers proposing 3-token blocks) and an NVFP4 checkpoint targeting Blackwell GPUs (vLLM, SGLang).
- Licence: Mistral Ecosystem Open-Weight License, a modified MIT allowing commercial use but requiring a separate licence to run a Model-as-a-Service business.
- Also served on the Mistral API as
mistral-large-4.