Mistral says its new model beats Chinese rivals on cyber, but shows no numbers

Update Oct 6: Euronews published an explainer on Mistral's new AI model and where it stands in the open AI race between the US and China. The supplied article details do not include any further information about the model. Euronews.com.

Mistral's CEO claims a new model tops Chinese models on cybersecurity. He named no rivals and no benchmark, so we read it as a pitch, not a result.

French AI company Mistral says its newest model outperforms Chinese models on cybersecurity. CEO Arthur Mensch made the claim onstage at a conference in Abu Dhabi, according to Reuters, shortly before the model was due to be unveiled. His words were that it is "above the Chinese models on certain aspects, including cyber," and that the idea Europe cannot compete "is not true."

Here is the catch: Mensch did not say which Chinese models he meant, nor what measure was used. That is the whole claim, and it is unverifiable as stated. A comparison against one older model on one narrow test is very different from beating the strongest Chinese systems across a broad evaluation, and the announcement as reported does not tell us which of those we are dealing with.

Mistral is spending on visibility in the US while pitching itself as a European alternative. Image: Unsplash — a man in a suit standing with his arms crossed

The context matters more than the claim. Mistral is three years old and, per Reuters, is seen as Europe's best contender in the AI race, though still dwarfed by Anthropic and OpenAI and squeezed by cheaper Chinese rivals. It raised €3 billion at a €21 billion valuation in September, and its last model launch was in April. A company in that position has every reason to pick a battlefield where it can claim a win, and cybersecurity is a sensible one to pick.

The more interesting part of Mensch's speech was commercial, not technical. He said independence from both the US and China is helping Mistral grow quickly in the Gulf and Asia-Pacific, because there is "an enormous need for alternative technology suppliers." That is the real product: a model that is good enough and politically neutral. Others are chasing the same niche, as we saw with Reflection's Beam, which also targets Chinese models on lower compute cost.

Our take: we will wait for the model card and independent evaluations before believing anything. If Mistral publishes the benchmark, the named competitors and the test conditions, this becomes news. If it does not, it stays a good line for a conference stage. Mistral's pitch of being a safe European option is credible on its own terms, but a claim of technical superiority needs numbers behind it, and so far we have none.

Further reading: Purchase Licensing Rights , opens new tab.

Opinion · What the press release leaves out

The claim is a sentence said on a stage, not a published result. Mensch did not say which Chinese models Mistral beat, or by what measure. Without a benchmark, a model name or a methodology, nobody can check it, and the word 'certain' does a lot of work in 'certain aspects'.

Also missing: the model's name, its size, its price, whether it is open-weight, and when customers can actually use it. The story is shaped around the 'Europe cannot compete' narrative, which suits a company that just raised money and wants Gulf clients.

Timeline

1. OCT 6 Euronews published an explainer on Mistral's new AI model and where it stands in the open AI race between the US and China. The supplied article details do not include any further information about the model. Euronews.com
2. OCT 6 Mistral says its new model beats Chinese rivals on cyber, but shows no numbers

Related articles from LaRebelión

South Korea bets $3.5 billion on its own frontier AI modelAI agents create fresh cybersecurity risks and liability dilemmasA&T Center Focuses on AI and National Defense for Security Month

Original sources: Reuters · Euronews.com

Produced with AI support and reviewed by the newsroom

Falcon · Signals analyst · Riyadh

“A CEO announcing a win on stage hours before the model launches is marketing until someone publishes the benchmark.”