Mistral Large 4

How it is built

Mistral Large 4 is a multimodal mixture of experts from Mistral AI with 1.05T parameters, 52B of them active for each token, and weights released on 29 October 2026.¹¹1Mistral AI documentation, Mistral Large 4 model page, version 26.10, read 30 October 2026. https://docs.mistral.ai/models/mistral-large-4-0 ²²2Mistral AI, "Le Chonk: Built for Coders, Trusted by Defenders", 29 October 2026, which releases the weights. https://mistral.ai/news/mistral-large-4-oss/ The count includes a vision encoder of 1.6B.¹¹

The language model has 72 layers. Only the first is dense, so 71 are mixture of experts layers.³³3Mistral AI, config.json in the official repository: num_hidden_layers 72, first_k_dense_replace 1, n_routed_experts 128, n_shared_experts 1, num_experts_per_tok 4, hidden_size 6144, moe_intermediate_size 6144, intermediate_size 24576, kv_lora_rank 512, max_position_embeddings 1048576. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/config.json. Arithmetic: 72 - 1 = 71 MoE layers; 6,144 / 24,576 = 1/4. Each holds 128 routed experts and 1 shared expert, and the router sends every token to 4 of the routed ones.³³

The hidden size is 6,144. Each routed expert is 6,144 wide, a quarter of the 24,576 in the dense layer.³³ Attention uses a compressed latent cache of rank 512 for keys and values.³³ The configuration supports 1,048,576 positions.³³

What is new

Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe.⁴⁴4Mistral AI, Mistral Large 4 model card, Hugging Face, read 30 October 2026. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B One set of weights covers instruction following and reasoning. It also handles coding and agentic work.⁴⁴

A request sets reasoning_effort to low, medium or high, and high is the default.⁴⁴ Mistral recommends a temperature of 1.0 and a top_p of 0.95.⁴⁴

On the Artificial Analysis Intelligence Index v5, Mistral reports 26, the highest among western open-weight models. Open models from China score higher on that index.⁴⁴ On the Coding Agent Index it scores 48%, level with GLM-5.3, while using 47% fewer tokens per task.⁴⁴

The licence is a modified MIT. Anyone may use the model, change it or sell it. A company that runs it as a hosted service needs a separate commercial licence from Mistral.⁴⁴⁵⁵5Mistral AI, licence file of the Mistral Large 4 repository, described on the model card. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE

Running it

The API opened as a public preview on 6 October 2026, before the weights came out.²²⁶⁶6Mistral AI, "Introducing Mistral Large 4", 6 October 2026, and the model page, read 30 October 2026, prices per million tokens. https://mistral.ai/news/mistral-large-4 The model is mistral-large-4, with a 1M context window.⁴⁴⁶⁶

Standard input costs $0.70 per million tokens, and output costs $2.10.⁶⁶ Cached input costs $0.07.⁶⁶ Output is 40% dearer than on Mistral Large 3, at $1.5.⁶⁶⁷⁷7Mistral AI documentation, Mistral Large 3 model card, version 25.12, output price $1.5 per million tokens, read 29 September 2026. https://docs.mistral.ai/models/model-cards/mistral-large-3-25-12. Arithmetic: 2.1 / 1.5 = 1.4.

At one byte per parameter the FP8 weights take about 1.05 TB (1.05T x 1 byte), before the cache.³³⁸⁸8Arithmetic: 1.05T x 1 byte = 1.05 TB, weights only. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B Mistral also publishes an NVFP4 checkpoint for Blackwell GPUs. It runs in vLLM and SGLang, and the card calls its performance the same.⁴⁴