French AI startup Mistral has revealed its latest large multimodal model: ML4, nicknamed “Le Chonk,” a one-trillion parameter model designed to push past both closed source giants and open models. Its release marks a strategic moment in Europe’s bid to stake a stronger claim in the evolving landscape of artificial intelligence.
What makes ML4 stand out
ML4 is built on the belief that closed models—which restrict access—and open source ones—often originating from China—aren’t the only paths forward. Initially, ML4 will be accessible only via a public API that includes guardrails to prevent misuse. Yet, in about three weeks, Mistral plans to release the full model weights so researchers and institutions can audit and utilize it beyond the API access, pending completion of safety evaluations.
Powered by a compute cluster of 4,000 NVIDIA GPUs—all work done in-house—ML4’s training regime was lean compared to rivals. Mistral claims its scale of infrastructure was “two to three times less” than major Chinese AI labs, and significantly less than closed-source providers, yet sufficient to achieve high performance. The model’s multimodal design is geared toward handling diverse input types and delivering value in specialized domains.
Focus areas, backers, strategy
Mistral is targeting sectors where precision and versatility matter most: cybersecurity, finance, chip design, among others. Chip design is particularly significant given that two of Mistral’s primary investors—ASML and Samsung—operate heavily in that space. These strategic alignments suggest ML4 is not only a technical project but also one with deep roots in Europe’s industrial landscape.
Though benchmarks for ML4 are still being finalized, Mistral is positioning it as likely the top open-weight model globally, especially outside China. Through ML4, the company hopes to exceed closed models in areas where its multimodal capabilities and targeted optimization matter most for enterprise use. Despite the wait for public weights, the API version already supports external use under protections.
Why this matters
This launch reflects growing concerns among enterprises about trust, safety, and transparency in AI. ML4’s planned weight release aims to satisfy demands for auditability without exposing the model to misuse prematurely. It also represents Europe doubling down on its vision for “a third way in AI”—neither fully closed nor unverified open, but a hybrid built with public oversight and industrial alignment.
It remains to be seen how ML4 performs on standard benchmarks once made fully available. For now, it signals that smaller labs—when aligned with strong industrial backers—can still build models that at least aspire to rival the scale and capability of tech industry behemoths.
Analytical angle: ML4 isn’t just another large model. It’s Europe laying down a marker: scale and openness can coexist with responsibility. The real test will come when the model’s weights—and its benchmarks—are public. If ML4 delivers, it could shift expectations around what open models must offer in terms of safety, performance, and industrial relevance.
