What Mistral Actually Built
Large 4 represents a significant jump from its predecessor. Mistral Large 3 scored just 9 points on the Artificial Analysis Intelligence Index, and the more recent Mistral Medium 3.5 reached 14; Large 4 is positioned well above both. The model uses a mixture-of-experts architecture, meaning it has roughly 1 trillion parameters in total but only activates about 49 billion of them for any given task, a design choice that makes it considerably more efficient to run than a dense model of the same overall size would be.
Mistral's own announcement describes the model as "state-of-the-art among open weights on cybersecurity, finance, and manufacturing," and says it achieves "performance competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe." That claim comes with an important caveat the company doesn't shy away from: on the same aggregated benchmarks, Large 4 still trails the leading closed models from Anthropic, OpenAI and Google, as well as several open-weight models out of China. Mistral's positioning is specifically about leading the US-and-Europe open-weight category, not claiming the overall frontier.
Where It Actually Performs Well
The company has put unusual emphasis on cybersecurity in its own announcement. On a benchmark that asks a model to reproduce a real vulnerability in open-source software and then patch it, Large 4 scored 82%, which Mistral says is the highest result of any model tested on that measure, open or closed. SiliconANGLE's coverage confirms the same 82% figure specifically on what's described as an "AA Cyber Index" score, putting Large 4 ahead of its open-source rivals in that category.
Mistral also highlights finance as a strength: the model reached 67% on FinWorkBench, a benchmark testing spreadsheet and accounting tasks in realistic finance scenarios, putting it roughly level with DeepSeek-V4-Pro and just ahead of GLM-5.3 at 65%, and nearly double what Mistral's own Medium 3.5 managed on the same test. The company also says third-party evaluators at vals.ai found Large 4 exceeding GPT-6 Astra on representative legal and financial tasks specifically, though this is Mistral's own characterization of that external evaluation rather than something vals.ai has been quoted on directly.
Beyond text, the model handles technical drawings, including converting them into CAD models, and performs strongly on semiconductor-related tests, a detail that points toward industrial and manufacturing customers as a deliberate target market, not just general enterprise users.
The Sovereignty Pitch
Much of Mistral's framing around this launch is explicitly political as well as technical. The company argues that companies and governments shouldn't have to depend on a provider that could restrict or cut off access to AI tools at any moment, a direct reference to concerns about relying on US or Chinese infrastructure for critical systems. Mistral says its customers specifically want guarantees that no foreign government can interrupt their access to the technology, and that running Large 4 end-to-end under European law, on European infrastructure, is central to that pitch.
This is also a funding story. Large 4 is described as the first major model to come out of the €3 billion Series D round Mistral closed in September, making this launch something of a proof point for that investment, Mistral's first big technical deliverable since raising the money.
What's Not Available Yet
A few details are still pending. The API preview is live now through Mistral Studio, but the actual downloadable model weights, which is what would let organizations run Large 4 entirely on their own infrastructure, aren't expected until the end of October. Mistral also hasn't yet published the specific license that will govern those weights when they do arrive; the model's current listing simply says "Open" without further terms specified. In the meantime, the company says it's testing a version of the model with reduced moderation and expanded cybersecurity capabilities under controlled conditions, specifically with cybersecurity leaders, vetted partners and state authorities, ahead of the broader public release.
