Mistral Large 3

How it is built

Mistral Large 3 activates 41B of its 675B parameters for each token.¹¹1Mistral AI documentation, Mistral Large 3 model card, version 25.12, released 2 December 2025, prices per million tokens, read 29 September 2026. https://docs.mistral.ai/models/model-cards/mistral-large-3-25-12. Arithmetic: 41 / 675 = 6.1%. Most of the total is the language model, at 673B with 39B active. The vision encoder adds 2.5B.²²2Mistral AI, Mistral Large 3 675B Instruct 2512 model card, Hugging Face. https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512

The language model has 61 layers. The first 3 are dense, and the other 58 are mixture of experts layers.³³3Mistral AI, params.json in the official repository: n_layers 61, first_k_dense_replace 3, num_experts 128, num_shared_experts 1, num_experts_per_tok 4, expert_hidden_dim 4096, hidden_dim 16384. https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512/blob/main/params.json. Arithmetic: 61 - 3 = 58 MoE layers; 4,096 / 16,384 = 1/4. Each of those holds 128 routed experts and 1 shared expert, and the router sends every token to 4 of the routed ones.³³

Mistral calls the design a "granular Mixture-of-Experts".¹¹ The experts are narrow. Each is 4,096 wide, a quarter of the 16,384 in the dense layers.³³ About 6.1% of the parameters work on any one token.¹¹

What is new

Mistral had not built a mixture of experts since the Mixtral series.⁴⁴4Mistral AI, "Introducing Mistral 3", 2 December 2025. https://mistral.ai/news/mistral-3 Large 3 was trained from scratch on 3,000 NVIDIA H200 GPUs.⁴⁴

The base model and the instruction-tuned model both came out under Apache 2.0.⁴⁴ At launch Large 3 ranked second among open non-reasoning models on LMArena.⁴⁴ Amazon Bedrock and Azure Foundry carry it as well as Mistral's own platform.⁴⁴

The model card lists its limits. Large 3 is "not a dedicated reasoning model", and on multimodal tasks it can trail models built for vision first.²²

Running it

The instruct weights are FP8, about 675 GB at one byte per parameter (675B x 1 byte).²²⁵⁵5Arithmetic: 675B x 1 byte = 675 GB, weights only, before the KV cache. https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512

The card puts the FP8 version on a single node of B200 or H200 GPUs.²² An NVFP4 checkpoint brings it down to a single node of H100 or A100 GPUs. A BF16 version is published too, with an EAGLE draft model for speculative decoding.²²

The vLLM command on the card sets a tensor-parallel size of 8 and a maximum model length of 262,144 tokens.²² For production, Mistral wants the temperature under 0.1.²²

On the API it is mistral-large-2512, at $0.5 per million input tokens and $1.5 per million output tokens, with a 256k context window.¹¹

Less than five months later Mistral released Mistral Medium 3.5, a dense 128B model. Its output costs five times as much, $7.5 per million tokens.⁶⁶6Mistral AI documentation, Mistral Medium 3.5 model card, released 28 April 2026, read 29 September 2026. https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04. Arithmetic: 7.5 / 1.5 = 5; 2 December 2025 to 28 April 2026 is almost five months.