Mistral Large 4 Is Here: 1T Parameters, 1M Context, Pricing, Benchmarks and Open Weights

How strong are the Mistral Large 4 benchmarks?

Mistral is making aggressive performance claims, but the safest way to read them is benchmark by benchmark. Some numbers come from third-party evaluation suites cited by Mistral; others are company-run or company-presented comparisons. None of these results should be confused with a Digital Pulse Brief hands-on test.

The most interesting result is not necessarily the largest number. Large 4 is being positioned as a model that combines coding, tool use, multimodal input and security work rather than optimizing for a single leaderboard. That is the same broad direction visible across the frontier-model market: users increasingly care about whether a model can operate across repositories, documents, images and tools in one workflow.

Why Mistral is emphasizing cybersecurity

Mistral says Large 4 ranks among the top five models on the Artificial Analysis Cyber Index and describes it as its strongest model for security work. The company is also doing something unusual before releasing the weights: it says cybersecurity leaders, vetted partners and state authorities will red-team the same model with reduced moderation and expanded cyber capabilities.

This matters because capable security models sit on a difficult line. Defensive teams may need a model to reproduce a vulnerability, analyze malware or test exploitability before they can patch a system. At the same time, those capabilities can be misused. Mistral argues that open weights and self-deployment can give legitimate defenders more control over model policy and availability, but organizations would still need their own access controls, logging, isolation and human review.

Why the open-weight release matters more than the nickname

“Le Chonk” is memorable, but the strategically important part of the launch is Mistral’s commitment to release the weights. If that happens on the announced schedule, organizations will be able to evaluate a frontier-scale Western model outside a vendor-only API environment and potentially run it on infrastructure they control.

Open weights do not automatically mean cheap or easy self-hosting. Large 4 has 1.05 trillion total parameters, and Mistral has not yet published all the architecture, post-training and deployment details it says will accompany the weights. Until those files and technical notes are released, concrete claims about GPU counts for inference, practical quantization quality or workstation deployment would be speculation.

The architecture does, however, help explain why the model is not equivalent to running a dense one-trillion-parameter network for each generated token. Its 49 billion active parameters are a key part of the efficiency story. The final deployment story will depend on expert routing, memory layout, weight precision, serving software, batching and the hardware used.

Trained in Europe on 3,800 Grace Blackwell GPUs

Mistral says Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the public preview is served on the same infrastructure. The company also says a substantial share of the training data was multilingual across more than 160 languages.

The infrastructure detail is central to Mistral’s sovereign-AI pitch. It is not merely selling model access; it is arguing that customers should be able to choose where inference happens, how models are deployed and, eventually, whether they run the weights themselves. Mistral’s regional-inference documentation separately offers EU and US endpoints for organizations that need to control the geography of inference processing, with a 10% regional upcharge over standard list pricing.

Mistral also says the reinforcement-learning run behind the preview is still active. At its current scale of roughly 3,000 GPUs, the company says a run generates about 33 billion tokens per day, leaving around 16 billion trainable completion tokens after filtering and masking. That makes the “preview” label meaningful: the model may still change before the weight release.

What Large 4 changes in the open-model race

The broader story is competition. For much of the open-weight frontier, Chinese model developers have set the pace on price-performance and downloadable capabilities. Mistral is presenting Large 4 as evidence that a European developer can compete at that level while combining open weights with its own cloud, regional infrastructure and enterprise customization stack.

Reuters reported that Mistral views the model as competitive with Chinese rivals in several areas, especially cybersecurity. That claim will become easier to assess once the weights are public and independent labs can run the same checkpoints under reproducible conditions.

For developers, the competitive pressure is useful regardless of which vendor wins a leaderboard. More strong open-weight options can reduce dependence on one API provider, create more deployment choices and push pricing down. For enterprises, the harder question is not simply “which model scores highest?” but which model meets requirements for accuracy, latency, data location, governance, integration and total cost.