Mistral Large 4 ships as a ~1T-parameter mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → (49B active) with a 1M-token context window, released in FP8 and NVFP4 variants under an open-weight license, and is positioned as particularly strong for cybersecurity and coding agents.
Key takeaways · AI-distilled
Mistral Large 4 uses two kinds of sparsity: a router picks 4 of 128 experts per token plus one always-on shared expert, and a DeepSeek Sparse attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → indexer picks up to 2,048 key positions per query on top of compressed Multi-head Latent Attention.
Mistral Large 4 scores 82% on a test of reproducing and then patching a real open-source vulnerability. Mistral reports several closed frontier models score near zero there because they refuse, yet Large 4 refuses harmful cyber prompts more than other open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models.
Reasoning in Mistral Large 4 cannot be turned off. Effort takes low, medium or high, high is the default, and per Hugging Face, passing reasoning_effort none still runs at high.
Hugging Face measured the 2-bit Q2_K quant on one 512 GB M3 Ultra at about 90 tok/s prefill and 15 tok/s generation. On four RTX6000 Pro cards, IQ2_XXS fits an 800K context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, but generation falls from 30 tok/s at short context to 3 tok/s at long context.
On Hugging Face's 8x H200 endpoint recipe with FP8 weights, a 1K-token prompt saw about 0.6 s to first token and roughly 65 tok/s, rising to 120 to 138 tok/s when the built-in DSpark draft head's guesses were accepted. The NVFP4 checkpoint in vLLM requires Blackwell GPUs.
Terms in this piece · Glossary
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
