Mistral's flagship model is now something companies can download and run on their own hardware. Hugging Face's launch post for Mistral Large 4 sets out what that means in practice: a model with about 1 trillion parameters, roughly 49B of them active for each token, a 1M-token context window, image input, and day-one support in Transformers, vLLM and llama.cpp. For security teams in particular, it is one of the few strong models whose behaviour they can fully control.
The design combines two kinds of sparsity. A router picks four of 128 experts for each token, plus one shared expert that always runs, and an attention indexer picks up to 2,048 earlier positions for each query rather than reading the whole context. A 1.6B Pixtral encoder handles images. On 29 October it ranked fifth among open-weight models on the Artificial Analysis Intelligence Index. Reported coding scores include 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.
Cybersecurity is the standout. The model solved 93% of 40 Cybench challenges taken from security competitions and scored 82% on a test that asks it to reproduce a real vulnerability in open-source software and then patch it. Mistral says several closed frontier models score near zero on that test because they refuse to try. The argument for open weights here is practical: defenders cannot afford a provider's filter blocking an incident response, or an API going down mid-incident. The post also says the model refuses harmful cyber prompts more often, on average, than any other open-weight model.
Running it is still a serious undertaking. A 2-bit quantised version fits on one Apple M3 Ultra with 512 GB of memory, where Hugging Face measured about 15 tokens a second of output. Hugging Face's ready-made hosted setup uses eight Nvidia H200 GPUs. Our review of Mistral Large 4 notes that independent tests still put it mid-pack among big open models, and it now undercuts Mistral Medium 3.5 on API price, which leaves the smaller model a harder sell.
Two catches deserve attention before you build on it. The licence is a modified MIT licence that limits model-as-a-service use, so a company that wants to sell hosted access needs a separate agreement. And reasoning cannot be turned off: the effort levels run from low to high, and asking for none falls back to the default, high, which adds latency and cost to simple tasks. Most of the benchmark numbers in the post come from Mistral, and the long-context speeds Hugging Face measured drop sharply, to about 3 tokens a second at very long contexts on four workstation GPUs.