Mistral Large 4: a 1 trillion-parameter model, weights to follow

Mistral AI has released Mistral Large 4, a new large multimodal model. According to TechCrunch, the model is nicknamed 'Le Chonk' inside the company because of its 1 trillion parameters. A parameter is a tunable value a model learns in training; as the number grows, a model generally becomes larger and more costly to run.

Is the model open to everyone right now?

Partly. The model is not yet open-weight; for now it can be used only through a guarded public endpoint. Mistral's VP Science Pierre Stock said the company plans to publish the weights in about three weeks, once safety testing is complete. Open weights mean the model's trained values can be downloaded and run on your own infrastructure, which makes auditing easier.

How was it trained, and what is the goal?

According to Stock, the model was trained entirely on Mistral's own compute with 4,000 Nvidia GPUs, which he said is far fewer than competitors use. Benchmark results have not yet been published. The company hopes the model will be the best among open-weight models, and thinks that thanks to focused training it could also beat closed models in specific areas. The highlighted use cases are cybersecurity, finance and chip design.

Chip design is tied to the work of the company's two main backers: ASML led its Series C, and Samsung led its Series D last month at a valuation of €21 billion.

What does it mean for enterprise teams in Türkiye?

For banks, healthcare organisations and software companies that want to run a model on their own servers for data-privacy reasons, an open-weight large model could be a valuable option. However, there are no independent benchmarks yet, so the 'better' claim should be read for now as the company's goal.

• Wait for benchmark results until the weights are published; plan a small trial with your own dataset.
• For data covered by KVKK, start calculating the cost and hardware requirements of running the model on your own infrastructure.
• Read the licence terms for commercial use once the weights are published.