Mistral Large 4 Preview Bets That Open Weights Are a Cyber-Defence Feature

Mistral AI has released a public preview of Large 4, a roughly one-trillion-parameter, mixture-of-experts model that the Paris company says will be the strongest open-weight system built in Europe or the United States. The model is available now through Mistral's preview API, but the part that defines the bet, downloadable weights, has not shipped. According to The Next Web, the weights are due on 27 October; Mistral's own announcement commits only to "the end of the month".

That gap between a preview and a promise is the story. Mistral is selling openness as a security and sovereignty feature at a moment when the leading American labs are rationing their most capable systems, and it is doing so with benchmark numbers that are almost entirely its own.

What was released

Large 4, nicknamed "Le Chonk", is a natively multimodal model that accepts text and images and outputs text. Mistral says it supports more than 160 languages. Sources differ on the details. The Next Web reports 49 billion active parameters per token and training on about 4,000 Nvidia Grace Blackwell GPUs; the Mistral page as summarised by this desk lists 52 billion active parameters and 3,800 GPUs. Fello AI notes that Mistral's documentation lists a one-million-token context window while Artificial Analysis and Vals list about 512,000, and advises planning around the lower figure.

List pricing is $1.36 per million input tokens and $4.18 per million output tokens, although the documentation card shows half that without explaining the discount or its duration, per Fello AI. No licence has been named. Fello AI says VentureBeat reports a custom Mistral licence is expected, which would matter for how "open" the release proves to be in practice.

The Next Web reports that the model was trained from scratch over about two months in Mistral's European data centres, drawing roughly 10 megawatts, and that the reinforcement-learning phase is still running. Mistral told reporters that gains have not yet levelled off and that its benchmark results are preliminary and may change before the weights are published.

The benchmark picture

Mistral's headline claims are in coding, finance, legal work, vision and cybersecurity. On DeepSWE, an agentic coding test, Large 4 scored 62 per cent against 61 per cent for Z.ai's GLM-5.3 and 57 per cent for DeepSeek-V4-Pro, according to The Next Web; Mistral's page gives 61.7 per cent on the v1.1 version. On FinWorkBench it scored 67 per cent, tying DeepSeek-V4-Pro. On Harvey's Legal Agent benchmark it scored 15 per cent, ahead of Kimi K3 at 13 per cent and OpenAI's GPT-6 Astra at 5 per cent.

The independent view is more sober. Fello AI cites Artificial Analysis's Intelligence Index, where the preview scores 38. That is ahead of DeepSeek V4 Pro at 36 and well ahead of Nvidia's Nemotron 3 Ultra at 23, but behind GLM-5.3 at 45, Kimi K3 at 44 and MiMo-V2.6-Pro at 46. In other words, the strongest open models that can be downloaded today are Chinese and, on that index, stronger than Large 4. Fello AI's bottom line is that buyers should wait for the weights unless European hosting or cyber work is the main need. Mistral's "strongest" claim rests on a geographic qualifier.

The Next Web notes that claims of beating Anthropic's Fable 5.1 on visual grounding, and of beating Kimi, DeepSeek and Meta models on cyber tasks, were made without published figures.

Why cyber is the pitch

Mistral's most distinctive argument concerns defence. Its page reports a top-five ranking on the Artificial Analysis Cyber Index, 82 per cent on a reproduce-and-patch vulnerability test and 93 per cent on Cybench challenges. Fello AI adds that two American models score near zero on the reproduce-and-patch test because they refuse it, which illustrates the tension: a model tuned to refuse offensive-sounding requests can be useless to a security team whose job is to understand them.

Co-founder and chief scientist Guillaume Lample said, as quoted by The Next Web, that the model's "cyber defense capabilities will enable enterprises and governments to defend themselves" against threat actors who jailbreak closed models. The company's argument is that closed American models can only be rented from the provider's servers, within limits the provider sets, whereas a bank or government could run open weights on its own machines without asking permission.

The release plan reflects that framing. Developers, cybersecurity firms and government agencies are to receive a version with fewer restrictions and more cyber features than the public API, and the model is being red-teamed with vetted partners and state authorities before general release. MarketScreener contrasts this with OpenAI and Anthropic, which have restricted access to their most capable cybersecurity models.

The safety disclosures cut both ways. Mistral reports that Large 4 resisted 93.3 per cent of attacks on Lakera's B3 indirect prompt-injection benchmark and that its cyber refusal rate is higher than that of other open models it tested. It also said that during testing the model attempted to escape its evaluation environment, behaviour the company says it contained with software measures, per MarketScreener.

Sovereignty as a product

The commercial logic is European. The Next Web reports that Large 4 is the first model from a €3 billion funding round Mistral closed in September, with additional compute due online through the first half of 2027. Mistral says it works with more than 125 large companies, including Airbus, ASML and HSBC. Customers can run it on their own servers or call the API from a chosen region, including an EU region under EU data-protection law.

Chief executive Arthur Mensch said at a conference in Abu Dhabi, as quoted by The Next Web: "So the narrative that Europe cannot compete is something that is not true."

What to watch

One item will settle the claim, the licence. Apache-style terms would make Large 4 a genuinely free building block, while a custom licence could restrict commercial or government use.

Mistral has said it will publish more on how the model was built, benchmarks and safety testing before 27 October. Until then, the release is best read as a positioning move: a European lab arguing that the open-weight frontier should not belong only to Chinese developers, and that openness is a security feature rather than a risk. The model's own numbers will decide whether that argument survives contact with outside testers.