What happened
French AI lab Mistral AI released Mistral Large 4 (ML4) on Tuesday. It's a new 1 trillion parameter multimodal model, reported by TechCrunch. Nicknamed "Le Chonk", the model is not yet open-weight.
For now, ML4 is only accessible via a public guardrail endpoint. Mistral plans to release the model weights in three weeks, following safety testing. The model was trained entirely on Mistral's own compute, using just 4,000 Nvidia GPUs.
How the room's reading it
The release is seen as Mistral's attempt to carve out a "third way" in AI — a European alternative to closed US models and open-weight Chinese ones. The conversation on X is focused on the three-week delay before the weights are released. Some developers are sceptical, seeing the "safety testing" as a hedge. Others see it as a responsible step, especially for the enterprise and institutional users Mistral is chasing.
The lab is positioning ML4 as a top performer for specific use cases like finance and chip design, which aligns with its key investors ASML and Samsung. With benchmarks still pending, most practitioners are in a wait-and-see mode, curious if the performance will justify the hype around its efficient training.
Sailfish's take
A 1T model is a serious piece of engineering, but for builders, it's just a headline until the weights are public. The three-week guardrail period is effectively a public beta. We'll be testing the endpoint for sure — particularly on the finance and cybersecurity tasks Mistral is flagging. But we wouldn't plan any production workloads around it until we can run the model ourselves.
The most interesting claim isn't the parameter count, it's the training on just 4,000 GPUs. That suggests serious architectural or data optimisation. If the performance holds up once the weights drop, that's the real story here. This is one to watch closely, but not to commit to.