Models Published: Oct 29, 2026

Mistral releases the weights of Large 4, a 1-trillion-parameter model with 298 GB 2-bit quants

What happened

Mistral published the weights of Mistral Large 4 on October 29, after a hosted preview that opened on October 6. It is a multimodal Mixture-of-Experts model with 1 trillion parameters, 128 routed experts with 4 active per token, about 52 billion active parameters and a 1M-token context window. It ships under the Mistral Ecosystem Open-Weight License, a modified MIT license that allows commercial use but requires a separate commercial license for companies operating a Model-as-a-Service business. Mistral recommends vLLM, also supports SGLang, transformers and llama.cpp, and offers an NVFP4 checkpoint for Blackwell GPUs; ggml-org's GGUF files run from 298 GB (IQ2_XXS) to 2.10 TB (BF16).

Why it matters locally: This is Mistral's largest model and you can host it yourself, but Hugging Face says only the 2-bit quants run on a single machine, an Apple Silicon Ultra with 512 GB of unified memory; the others need distributed inference.

Three ways to read it

For the market

Mistral reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task, and calls it the top Western open-weight model on the Intelligence Index v5 at 26, while conceding Chinese open models score higher overall. The Model-as-a-Service clause suggests Mistral wants company use to spread freely while keeping hosting rivals inside a paid license.

In plain words

Mistral, a French AI company, has made its biggest model free to download. Think of it as a very smart employee who needs a whole office floor: most companies will still reach it online, because even the smallest version needs a computer with far more memory than a normal PC.

For practitioners

Plan on 298–381 GB for the 2-bit GGUFs (IQ2_XXS, Q2_K), 635 GB for Q4_K_M and 1.12 TB for Q8_0; ggml-org says the 2-bit files fit in 512 GB of unified memory. On NVIDIA hardware, the NVFP4 checkpoint is about 610 GB of weights, targets B200, B300, GB200 and GB300, and needs --kv-cache-dtype fp8_ds_mla in vLLM, with Mistral's sample commands using 8 GPUs. Reasoning cannot be turned off in the GGUF builds, so test low effort for latency, and have legal read the license if you resell access.