Mistral AI, a French AI lab, has put out a multimodal model with a trillion parameters. The company is positioning itself as a middle ground in the AI industry, sitting between two groups: closed models that can be switched off and open-source models that frequently come from China.
The model, named Mistral Large 4 (ML4) and nicknamed Le Chonk, was announced Tuesday. It follows a phrase from French President Macron, who has described a “third way in AI.” Mistral is positioning ML4 as an alternative to both American and Chinese rivals.
The Chonk’s Weight Problem
The ML4 weights are not available for public access at this point. Instead, they can currently be reached via a public guardrail endpoint. Once safety testing is finished, Mistral has said it will release its weights within three weeks.
VP Science Pierre Stock told TechCrunch that the company will work with trusted partners and governments before the weights are released. “In the meantime, we’ll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks,” he said.
Security and the Open-Weight Trade-Off
The security worries have risen among Mistral’s key audience — businesses and institutions — even as Stock pointed out that an open-weight design is simpler to check.
Mistral wants to keep two things at once. One side lets a model be turned off. The other side lets it be opened up for close inspection. Each choice gives something up. Closed models can be unplugged, while open models can be studied closely.
Training on 4,000 Nvidia GPUs
According to Stock, Mistral built ML4 using its own hardware, with just 4,000 Nvidia GPUs involved in the training process. That number represents two to three times fewer GPUs than what Chinese rivals use, and it also falls short of the count employed by closed-source competitors.
Mistral’s efficiency claim stands out. The company contends it can train a top-tier model using far less in resources than its competitors require.
Benchmark Results Pending
Mistral is still waiting on benchmark results for its ML4 model, which the company believes will rank at the top of open-weight models, particularly outside of China. It also argues that the model may exceed the performance of closed models in specific domains that matter to its customers, where multimodal capabilities can add value.
Stock pointed to several optimized use cases:
| Use Case | Description |
|---|---|
| Cybersecurity | Core customer need |
| Finance | Core customer need |
| Chip design | Key to ASML and Samsung backing |
The Backers Behind Le Chonk
Two of Mistral’s main backers are Dutch giant ASML, which led its Series C, and Samsung, which led its Series D last month at a €21 billion valuation, about $24.39 billion.
The company attempted to communicate during that funding round that hosting Chinese models was not a pivot toward becoming merely an inference provider. It believes that with Le Chonk on its side, it can still be regarded as a frontier lab.
What This Means for Open Source
Stock presented the approach as a joint effort, with trusted partners and governments working together. The weights will be released after testing is finished, not ahead of time.
ASML led Mistral’s Series C, while Samsung has also invested heavily. Both stand to gain from a model that performs well in chip design.
The Verdict on Le Chonk
Mistral is presented as a lab capable of training a top-tier model using fewer resources than its rivals, yet it holds its weights back under guard until testing is complete.
There is a creation already made. Its examination now commences. What remains to be seen is whether the numbers that govern it come without any concealment.
A trillion parameters, a French nickname, and a lab aiming to offer a third path in AI: that combination makes up the idea behind Le Chonk, which has become the name for the scope of its goals.
Source material: “Mistral’s new 1T model aims to leapfrog closed and open rivals,” TechCrunch.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.

