Mistral Large 4

About

Mistral Large 4 (mistralai/Mistral-Large-4-1T-A52B) is Mistral AI's largest and most capable model to date — a 1-trillion-parameter, natively multimodal granular Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers). It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters and released as open weights on October 29, 2026 ("Le Chonk"), after a public preview of the flagship on October 6, 2026.

The text backbone is a 72-layer decoder-only MoE transformer: 128 routed experts with 4 active per token plus 1 shared expert (first layer dense), multi-head latent attention (MLA) with sparse indexed attention layers — a lightning-indexer-style learnable attention index (index head dim 128, 32 index heads, top-k 2048) — and RoPE with YaRN scaling supporting a native 1,000,000-token (1,048,576) context window. Vision comes from a Pixtral-style vision encoder (30 layers, 14px patches, 1540px base image size) trained for variable image sizes and aspect ratios: the model accepts text + image input and produces text output. It is natively fluent in 160+ languages, including every official language of the European Union.

The checkpoint ships FP8-quantized weights (dynamic activation scheme; embeddings, vision tower, projector, lm_head and KV/index projections stay BF16/F32) plus an embedded DSpark speculative decoding draft head: 3 dense MLA draft layers in BF16 with a recurrent Markov head proposing blocks of 3 tokens, measured up to 1.9x higher throughput without changing target outputs. A separate NVFP4 checkpoint targets Blackwell GPUs.

Mistral Large 4 unifies instruction following, reasoning, coding, and agentic workflows in one set of weights: three reasoning-effort levels (low/medium/high, default high), native function calling, JSON output, strict JSON schema enforcement, and strong system-prompt adherence. Performance highlights (Mistral/card, Oct 2026): best open-weight model for coding — Artificial Analysis Coding Agent Index 48% (effectively tied with GLM-5.3 at 49%, with 47% fewer tokens per task), Terminal-Bench 4.0 41% (open-weight lead), Cognition FrontierCode 2.0 19.7% (lead), and the highest score among western open-weight models on the Artificial Analysis Intelligence Index v5 (26, vs. 32 for GLM-5.3). It is SOTA among open weights for cybersecurity (AA Cyber Index 51%, 96.6% average refusal of malicious cyber requests), strong for finance (FinWorkBench/Finch 73.3%) and law, web browsing (BrowseComp 84.5%), and agent security (B3 attack resistance 97.1%).

Deployed with vLLM (recommended), SGLang, transformers (AutoModelForMultimodalLM), llama.cpp (GGUF); fine-tuned with Axolotl, Unsloth, TRL. Served on the Mistral API as mistral-large-4 ($0.70 input / $0.07 cached input / $2.10 output per M tokens) and via OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints. Licensed under the Mistral Ecosystem Open-Weight License — a modified MIT license allowing commercial and non-commercial use; operating a Model-as-a-Service requires a separate commercial license. A full technical report is announced as upcoming.

Training Data Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe; continued reinforcement learning after the October 6 public preview. No training dataset names disclosed on the model card.