Mistral Large 4 (Chonk) to compete with American and Chinese tech giants

Nickname aside, “Le Chonk” is a name that makes you smile. Behind it, Mistral Large 4 is the most ambitious model the Paris-based start-up has presented to date, and its result is both flattering and modest.

On October 6, Artificial Analysis, an independent organisation that benchmarks artificial intelligence models, gave it 38 points on its intelligence index, version 4.3.2. That makes it the highest-rated model outside China among those Mistral describes as “open-weight”, meaning that their parameters can be downloaded and run by the customer.

That achievement remains relative. On the same chart, Large 4 sits 18th out of 25 models, level with OpenAI’s GPT-6 Luna. Anthropic’s Claude Opus 5.5 is credited with 58. Should we see this as a failure? I think the question is badly framed, and that the answer sheds light on a strategy that reaches well beyond the labs.

A ranking that calls for some caution

Let us first be clear about what this figure measures. The Artificial Analysis index aggregates ten evaluations: office-type tasks, programming, scientific reasoning, knowledge recall. It is revised frequently, and the organisation warns that scores from different versions cannot be compared with one another. A 38 today is therefore not comparable with a 38 from last year.

A second caveat: at this stage, Large 4 is a pre-release version. Mistral describes it as an open-weight model, but those weights had not been published at the time of writing. Artificial Analysis therefore classifies it, provisionally, among proprietary models. The organisation expects it to rank among the top three open models on its cybersecurity index once the weights are released. The ranking could also move if Mistral continues training and the final model is re-evaluated.

Finally, there is the cost. According to Artificial Analysis, running one task from the index costs an average of $1.13 with Large 4, against $0.03 for Large 3. The organisation judges this cost to be more than four times higher than that of open models of comparable intelligence. The 18-point gain over Large 3 therefore comes at a price.

Two races no longer run on the same scale

These two facts, a model leading its category and a mediocre rank in the global table, coexist because Mistral is no longer quite competing in the same contest as its two big American rivals.

In 2026, OpenAI closed a funding round of around $122 billion in March, at a valuation of $852 billion. Anthropic closed a $65 billion Series H in May, valuing the company at $965 billion. These sums finance computing power and training runs that are out of reach for a European player of this size. Trying to catch up with frontier models on their own ground would, on a comparable budget, amount to a lost bet.

Mistral seems to have drawn its conclusion. Its offer is no longer simply “the best model” but “the best model you can control”: knowing where it runs, under which legal regime, with what service guarantees, and being able to operate it without depending on a third party’s goodwill. The argument appeals above all to players for whom an interruption would be costly: government departments, banks, manufacturers, infrastructure operators.

Hosting in-house is not enough

The Bercy episode in June illustrates the limits of a purely technical approach to sovereignty. The French Treasury Directorate-General, part of the finance ministry, had deployed an assistant called HéphAIstos in early June to around a hundred officials, built on Qwen, Alibaba’s model. Several users reported answers deemed skewed on China-related topics. The experiment was halted on June 23, and a Mistral model was installed the following Wednesday, according to information reported by AFP and Le Monde.

Bercy stressed two points: the assistant ran offline, with no Internet access and no identified backdoor, and the test was never meant to last. The ministry thus ruled out any risk of data leakage. But the essential point lies elsewhere. A model run on one’s own servers transmits nothing externally, yet it can still carry the leanings of its training in its parameters. A responsible-AI specialist interviewed by AFP noted that biases are inherent to any model, but that some can be introduced deliberately.

A study by the German lab Aleph Alpha, covering 967 sensitive topics, points in the same direction: the Chinese models tested (Qwen, DeepSeek, Kimi) produced balanced answers in only 17 to 41% of cases. Claude Sonnet 5 reaches 70% and Mistral Small 92%. These results should nonetheless be read with caution, since Aleph Alpha is itself a European competitor and scored the answers with its own automated evaluation system.

The bet on the harness, the infrastructure and the law

The same logic explains a choice that surprised many. Since August, Mistral has offered its customers GLM-5.3, an open-weight model from the Chinese lab Z.ai, hosted on its own servers. On September 21, it made it the default option in the web interface of its coding assistant, Vibe Code. According to Reuters, the model is offered without modification. Engineers have nonetheless pointed out on social media that GLM-5.3 remains very close to the previous version, GLM-5.2, already available at Mistral.

At first glance, the inconsistency is glaring: how can one champion a European AI and distribute a Chinese model? Mistral replies that this open model, like those that follow, will run on the same infrastructure, with the same regional controls and the same service commitments as its own models. In other words, what the company sells is not just a model but what surrounds it: the “harness”, meaning the set of tools, safeguards and orchestration that make it usable in a business setting, along with hosting and compliance with European law.

This position has its critics. They see it as a renunciation: if the start-up’s flagship model is no longer the central argument, what remains of the original promise? The stakes are not minor, because Mistral’s credibility rests on its ability to train its own models. Defenders of the strategy answer that it is pragmatic: no customer chooses a supplier for the nationality of its parameters, but for the guarantee of not being caught out the day a contract, a law or an export control changes.

The figures to keep in mind:

• 38: Mistral Large 4 Preview’s score on the Artificial Analysis index (v4.3.2), against 58 for Claude Opus 5.5.
• 18th out of 25: its rank on the published chart, level with GPT-6 Luna.
• $1.13: average cost per index task, against $0.03 for Large 3.
• $122 billion and $65 billion: 2026 funding rounds of OpenAI and Anthropic.
• June 23: date the Qwen test was halted at the Treasury Directorate-General.
• More than one million public servants: target of the “Notre IA” plan, with a Mistral-powered assistant, for an initial budget of €700,000.

The question facing executives, CIOs and public-sector buyers

That leaves the question this reasoning poses to every organisation: if your American supplier cuts you off tomorrow, what stops working in your business?

The scenario is far from abstract. In June, Anthropic suspended access to two of its most advanced models in order to comply with export controls imposed by the US Department of Commerce, before restoring it on July 1, once those controls were lifted. The episode, which lasted only a few weeks, is a reminder that a model consumed through an API remains subject to decisions taken outside the customer company. That observation also applies to Moroccan and African organisations now building their tools on American models.

Three avenues emerge in response. Map the processes that rely on an external model, since not all of them are critical. Plan for a fallback model, possibly a less capable one, able to provide a minimum level of service. And demand transparency from suppliers on hosting, applicable jurisdiction and how the model behaves on sensitive topics.

Sovereignty still to be proven

It would be unwise to conclude that Mistral’s strategy has proved itself. The flagship model remains mid-table and expensive to run. Its final version, and its licence, are not yet known. The contract with the civil service shows that the state has confidence in it, but it has yet to be tested by large-scale use.

The real question is whether control is a lasting commercial argument, or an asset that will fade if American models one day offer comparable guarantees. In the meantime, companies have a decision to make, one that depends neither on a ranking nor on a nickname: choosing what they are prepared to delegate, and to whom.

FAQ

Is Mistral Large 4 really an open model?

Mistral presents it as such, but its weights had not yet been published at the time of writing. Artificial Analysis therefore classifies it provisionally as proprietary.

Why does Mistral offer a Chinese model?

Because its commercial argument rests on infrastructure, regional controls and the European legal framework, more than on the origin of the model. The company says it applies the same service commitments to GLM-5.3 as to its own models.

What happened at Bercy with Qwen?

On June 23, the Treasury Directorate-General halted a test of the HéphAIstos assistant, built on Qwen, after reports of answers deemed skewed on China-related topics. Bercy stresses that the tool ran offline.

What does this change for a company?

It encourages companies to map their dependence on AI suppliers, plan for a fallback model and demand transparency on hosting and on how models behave.