Mistral Large 4 goes open weight: good news, but I won't be running it

At the end of October, Mistral published the weights of Large 4, a model of roughly a trillion parameters. For Europe it is real progress. For someone serving models on an L4, the numbers add up fast, and the licence has stepped back compared with Large 3.

At the end of October, Mistral put the weights of Mistral Large 4 on Hugging Face. Roughly a trillion parameters, a mixture-of-experts architecture, text and image input, and an announced context window of one million tokens. It is presented as the best open-weight model in Europe and the United States, with Mistral acknowledging along the way that Chinese models such as DeepSeek, Kimi or GLM stay ahead in several areas.

I read a lot of "open European AI is back" this week. I am genuinely happy for Mistral. But as someone who runs models on his own infrastructure, mostly I reached for a calculator.

The calculation nobody puts in the headlines

The memory needed to load a model is the number of parameters times the size of each parameter.

  • In 16-bit (2 bytes): about 2 TB of weights.
  • In FP8 (1 byte): about 1 TB.
  • In 4-bit (half a byte): about 500 GB.

My LLM runs on an NVIDIA L4, with 24 GB of memory. Just loading the FP8 version would take about forty of them, before counting the KV cache. With 80 GB H100s you need 13 for the weights alone, so in practice two 8-GPU servers. According to the press, the quantised versions run on 4 to 8 B200 or B300 GPUs, NVIDIA's current top of the range.

The answer I will get: it is a mixture-of-experts, only a little over 50 billion parameters are active per token (that is what the repository name says, Mistral-Large-4-1T-A52B). True, and it cuts the compute per token. Memory, though, does not go down: every expert has to be loaded, since you cannot know in advance which ones the router will pick.

"Open", but for whom?

Open weights serve a specific purpose. A public administration, a large company or a hosting provider can run the model on its own hardware, audit it, specialise it, without depending on an American or Chinese API. For European digital sovereignty that matters, and I am not downplaying it.

For an independent developer or a small team, a model this size is used through an API, exactly like a closed model. The real gain will come from the smaller versions, and from the licence they ship under.

The licence stepped back

Large 3 was released under Apache 2.0. Large 4 comes under a custom licence, the "Mistral Ecosystem Open-Weight license", described as a modified MIT. The main restriction: a company that resells access to the model as a service has to obtain a commercial licence from Mistral.

I understand the logic. Training a model on nearly 4,000 GPUs costs a fortune, and Mistral does not want hosting providers reselling its work without paying anything back. But it means the players able to serve this model at scale, the ones with the GPUs, are exactly the ones who need permission. It moves away from open source in the strict sense, and that is a real change from the previous generation.

What I am actually waiting for

Mistral announces distilled versions "in the coming months", including one that would run on a laptop, with no date given. For me, that is the interesting part of the announcement. A good 7 to 30 billion parameter model, distilled from Large 4, under a clear licence: that I can serve with vLLM on my L4, measure, and plug into my RAG.

Until then, Large 4 as open weights is good news for the ecosystem, and one more line in the list of models I will not run myself. Both are true at the same time, and I would like announcements to say so just as clearly.

Sources