Mistral Large 4: what does the European artificial intelligence giant have to offer?
The French model, with a trillion parameters, promises competitive performance in software, finance and cyber security.
A trillion parameters and a nickname taken from those chubby cats in memes: ‘Le Chonk’. Mistral Large 4, or ‘Le Chonk’ as it’s known, was unveiled yesterday, 6 October, by the French company and is set to make its way into businesses, particularly European ones. For now, the weights are not publicly available: Large 4 is being previewed via API. Mistral has announced its release by the end of October. It is therefore an announced open-weight model, with public distribution yet to be finalised. Let’s take a look at how it’s put together and what its strengths are compared to US Big Tech. Spoiler: it’s designed to be the European alternative to closed American models and open Chinese ones.
First and foremost, the dimensions.
Parameters are the numerical variables learnt during training. Large 4 has around one trillion of them, but activates 49 billion for each token – each fragment of text being processed. The architecture, known as mixture of experts, distributes the work amongst the network’s components and reduces the computational load compared to a system that utilises all of them at every step. Memory requirements remain high: the model is unlikely to fit on an office laptop. It is also difficult at this stage to compare it with the Big Tech firms. Anthropic and Google do not disclose the sizes of their flagship models, only the benchmark results. There is therefore no basis for comparing them to the one trillion parameters claimed by Mistral. Furthermore, Large 4 activates around 49 billion per token: the overall size describes the architecture, but is not sufficient to determine which model is more capable.
What can it do?
As stated on the website, Large 4 processes images alongside text, a capability that is useful for interpreting charts and documents in which the layout contributes to the meaning. Mistral claims to have been trained in over 160 languages, including all the official languages of the European Union. This offers an interesting prospect for those operating across multiple countries, although it does not guarantee equivalent performance in every language. The company also claims competitive results in programming, finance, law and cybersecurity. The model could assist an analyst, correct code or help examine a vulnerability. Finally, Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centres, which are also hosting the public preview. The company thus controls a larger part of the supply chain, whilst still utilising US technology.
Why it is difficult to compare them to Big Tech
An advantage in finance or cyber security does not imply an advantage in writing, maths or programming. To compare results, the same test, the same tools and a comparable computational budget are required. On paper, then, the distribution of files containing the learned parameters is certainly of interest to European companies; it should enable them to install and adapt the model within their own infrastructure, in accordance with the licence. For the time being, however, the weights are not publicly available. It is therefore difficult to assess.
Mistral is therefore trialling the model with verified partners and public authorities. The figures are due to be announced by the end of October. The licence and actual costs will clarify how feasible it will be to adopt it. For businesses, the litmus test will be their own documents and processes: that is where a thousand billion parameters will have to prove their worth.
