French artificial intelligence company Mistral AI officially announced its largest model to date, Mistral Large 4, on October 6, 2026. Internally developed under the codename Le Chonk, the system marks the European startup's entry into the trillion-parameter domain. With this release, the Paris-based laboratory emphasizes its ambition to rival frontier systems globally while staying committed to distributing open weights.
From an architectural perspective, Mistral Large 4 relies on a mixture-of-experts design. Out of approximately one trillion total parameters, only 49 billion are active per token, rising to around 52 billion when counting embeddings and the output head. This sparse structure allows the network to store an expansive knowledge base while keeping inference compute requirements comparable to much smaller dense models.
The company placed particular emphasis on compute efficiency during pretraining. Mistral trained the system on an infrastructure cluster of just 3,800 to 4,000 Nvidia Grace-Blackwell GPUs. Leadership framed this comparatively lean hardware footprint as evidence that European teams can build competitive flagship models using a fraction of the compute typically mobilized by US or Chinese frontier competitors.
In initial evaluations, Mistral Large 4 demonstrated native multimodal capabilities and marked strengths in software security workflows. On the AA Cyber Index benchmark, the model registered an 82 percent success rate in remediating software vulnerabilities. It also logged strong scores on the Dense200 visual grounding test, reflecting enhanced accuracy when aligning complex visual input with textual representations.
The model is currently accessible in public preview through the Mistral Cloud API, offering developers two distinct reasoning settings labeled none and high. Open-source integrations, such as the llm-mistral plugin, became available alongside the launch. Mistral scheduled the full release of open model weights for late October 2026, reinforcing its track record of supplying accessible foundation models to researchers and enterprise developers.
The rollout of Mistral Large 4 delivers a notable boost to Europe's competitive standing in artificial intelligence. By pairing massive parameter capacity and lean execution with open weights, Mistral challenges closed-source providers and illustrates how architectural refinements can offset raw hardware scale.