Mistral Large 4: Europe’s free 1-trillion-parameter AI model, explained
French AI company Mistral has released the open weights of Large 4, a 1-trillion-parameter model aimed at coding and cybersecurity. Here's what it is, how it performs, and who can actually run it.
French AI company Mistral has released one of the biggest open AI models ever. Mistral Large 4 has 1 trillion parameters, and its “weights” (the files that make up the model) are now free to download. Mistral says it’s built especially for coding and cybersecurity.
What does “open weights” mean?
Most top AI models, like those behind ChatGPT or Gemini, are closed: you can only use them through the company’s app or API. An open-weight model lets anyone download the model and run it on their own computers or servers. That means:
- Companies and governments can keep their data in-house.
- No one can suddenly cut off access. As Mistral’s VP of Science put it, access “can’t be cut off by definition because the model is open-weighted.”
- Developers can customize it for their own needs.
Key facts
- Size: 1 trillion parameters, with 52 billion active for each task. Only part of the model works at a time, which makes it more efficient.
- Inputs: reads text and images; writes text and code.
- Reasoning modes: low, medium and high, so users can trade speed for depth.
- Where to get it: weights on Hugging Face, plus Mistral’s API. API pricing hasn’t been published yet.
- Training: on 3,800 Nvidia Grace Blackwell GPUs in Europe, with extra training on Mistral’s own computers in France.
How good is it?
- Coding: 41% on Terminal-Bench 4.0, up from 28% in its preview, just ahead of China’s GLM-5.3 (40%). Mistral says it matches GLM-5.3 on a popular coding-agent test while using 47% fewer tokens per task.
- Cyber defense: 79% on CyberGym-E2E, ahead of DeepSeek V4.1 Flash (76%).
These are Mistral’s reported results. Independent testing will show how it holds up in real use.
Mistral Large 4 is Europe’s answer to giant closed AI models: powerful, downloadable, and under your own control.
Why cybersecurity?
Mistral calls it “sovereign cyber defense” for companies and governments. Security teams often can’t send sensitive data to outside AI services. With open weights, they can run a powerful model inside their own systems to find vulnerabilities and analyze threats.
Can you run it at home?
Not really. Even the compressed (quantised) version needs four Nvidia B200 or B300 GPUs, data center hardware worth hundreds of thousands of dollars. Mistral plans smaller versions in the coming months, including one that can run on a laptop, but hasn’t given dates.
Why it matters
- Open AI is catching up: open models from Europe and China are getting close to the best closed models.
- Europe’s AI independence: Mistral is Europe’s leading AI company, and governments want AI they control.
- More choice for developers: another strong option alongside models from Meta, DeepSeek, Qwen and others.
Frequently asked questions
Is Mistral Large 4 free?
The weights are free to download from Hugging Face under Mistral’s license. Running it requires powerful hardware, and API use will be paid.
How many parameters does Mistral Large 4 have?
1 trillion in total, with 52 billion active for each task.
What is Mistral Large 4 good at?
Coding, agent tasks and cybersecurity, according to Mistral’s benchmarks.
Can I run Mistral Large 4 on my computer?
No, it needs data center GPUs. Smaller laptop-friendly versions are planned.
