With Large 4, Mistral Tries to Put Europe Back in the A.I. Race
Trending Topics Jakob Steinschaden ● Covered by 5 sources
Mistral just launched Large 4, a huge open model built to catch up with China and the US. It’s a Europe-vs-the-world bet, but the proof won’t come until the weights are out.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
After months of silence, Mistral came back swinging. The French company has unveiled Mistral Large 4, or ML4 internally, and it wants the model to read as proof that Europe can still matter in serious A.I. The message is blunt: we’re not done here.
Large 4 is a mixture-of-experts model with one trillion parameters, though only 49 billion are active at any one time. Mistral says it can keep pace with the strongest closed models while taking on open models that are three times larger. For now, that is mostly Mistral talking about itself. Independent checks on the open weights do not exist yet, and they won’t until the weights are released at the end of October.
The company is starting with an API preview priced at $1.36 per million input tokens and $4.18 per million output tokens, with downloads and self-hosting to follow later. That order matters. Mistral is aiming first at companies running models in private clouds or their own data centers, not at people hoping to run something this size on a laptop. It is simply too big for that.
The pitch to enterprise customers is very specific. Mistral says Large 4 is tuned for coding, agents, cybersecurity, domain-heavy work in manufacturing, finance and geospatial analysis, plus visual grounding in images such as satellite and aerial photos. It is multimodal, handles text and images, and was trained on more than 160 languages. The company also says it is especially strong on cybersecurity, where customers want a European model that can find flaws and defend against attacks without disappearing on them later.
Mistral’s confidence rests on its own benchmarks. It says Large 4 beats Z.ai’s GLM 5.2 in-house, edges out GLM-5.3 on DeepSWE, ties DeepSeek V4 Pro on FinWorkBench, and tops the chart on a vulnerability-reproduction test where closed models refuse to play along. But the picture is messier once outside comparisons enter the room: in a blind coding test, Large 4 comes second to Claude Opus 5. And Mistral is careful not to claim a clean win over China’s best overall.
That caution is part of the story. Europe keeps getting sold a familiar dream: sovereign A.I., industrial strength, local control, no foreign choke points. Mistral is one of the few companies that can credibly sell that dream, which is exactly why every launch gets read like a continent-wide exam result. The exam, annoyingly, is still ungraded.
My take — AI-written commentary, not fact-checked reporting
Mistral is doing the right thing by leaning hard into open weights and European control, because dependence is the real bug here, not benchmark theater. But the industry’s favorite trick is still alive: declare victory first, release the evidence later, and hope nobody checks the homework too closely.
Read more about this at: Trending Topics