How to use Mistral Large 4: three ways in, and the hardware catch

1. Through the Mistral API. This is the simplest path if you build software or use coding tools that accept an API key. You send requests to Mistral’s servers and get answers back, with no special hardware. The catch: Mistral hasn’t published final API prices for this release yet, so check its pricing page before you wire it into anything important.
2. Through Vibe. Mistral says the model is available to some Vibe users. If you already use it, look for Large 4 in the model options. Not every account has it yet.
3. Download the weights from Hugging Face. The weights are the model’s learned settings, the file that actually is the AI. They’re public, so you can run the model on your own machines and nobody can switch it off. This is the route for companies and governments with their own server rooms.

Not on a home computer. Large 4 has about 1 trillion parameters, the numeric settings a model uses to make its predictions. It doesn’t use all of them at once: it switches on 52 billion for each task, which keeps it faster than its size suggests. But all of them still have to sit in memory.

Mistral is releasing a quantized version, meaning the numbers are stored with less precision so the files shrink. That version runs on four Nvidia B200 or B300 GPUs. The full-precision model needs more than that. If you don’t own that kind of hardware, renting GPU servers from a cloud provider is the only way to host it yourself.

For everyone else, the real answer is to wait. Mistral says smaller, distilled versions will arrive in the coming months, including one that runs on a laptop. Distilling means training a smaller model to copy a big one’s behavior, so you keep much of the skill in a far smaller package.

For coding agents, it’s now among the best you can download. Mistral spent three extra weeks training the model after its Oct. 6 preview, using customer feedback, and the jump shows on Terminal-Bench 4.0, a test of AI agents working in a computer terminal. Large 4 went from 28% at preview to 41%, edging past China’s GLM-5.3 at 40%.

The overall picture is closer. On a combined index across three benchmarks, Large 4 scores 48, one point behind GLM-5.3 at 49 and level with Codex GPT-6.1 Terra. Mistral also says it matches GLM-5.3 on a coding-agent ranking while using 47% fewer tokens (the chunks of text a model reads and writes) per task, which matters because tokens are what you pay for.

In security work, it scored 79% on CyberGym-E2E, a test where a model must find, reproduce and patch real memory bugs in open-source C and C++ code. That matched MiMo-V2.6-Pro and beat DeepSeek V4.1 Flash at 76%. Mistral says it’s also more likely than other open models to refuse malicious hacking requests. As always, treat benchmark scores as a guide, not a guarantee.

Early testers asked for more control over how long the model thinks before answering. Mistral responded with low and medium reasoning modes alongside the existing high mode, and says high mode is now 50% more efficient. If you use the API for quick tasks, a lower mode should mean faster and cheaper replies.

The model reads both text and images and writes text and code. Some customers already use it for automated penetration testing, where AI agents probe a company’s own apps for weak spots and write up what they find.

Mistral’s pitch is control. Pierre Stock, its VP of science, compared AI to electricity and said nobody should depend on a supplier who can cut them off overnight. Once weights are online, he argued, access can’t be withdrawn. That’s a direct contrast with closed models from OpenAI, Anthropic and Google, which can be switched off remotely.

Critics point out the flip side. A security expert at Sophos called it clearly dual-use technology, since anything free to download can be misused too. Mistral also built Large 4 on a far smaller budget than its rivals: CEO Arthur Mensch says the company raised roughly $6 billion in three years.