Mistral Large 4: performance, security and the reality of sovereign open weights

On October 6, 2026, Mistral previewed Mistral Large 4, also known as "le Chonk." This is their latest and largest multimodal model. It uses a Mixture of Experts (MoE) design, scaling to 1 trillion total parameters while activating 52 billion parameters per token to keep inference efficient. It processes both text and images. Since the model is currently in preview and training is still ongoing, these performance numbers may still shift before the final release.

Right now, users can access the model through Mistral Studio. Mistral plans to release the model weights for open download at the end of October 2026, allowing for local deployments.

Infrastructure and Runtime Pricing

Mistral trained ML4 on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers, supported by a recent €3 billion funding round.

For those using the API, the pricing is straightforward:

• Input tokens: $1.36 per million tokens
• Output tokens: $4.18 per million tokens

Agentic Operations and Engineering Benchmarks

ML4 is built for multi-step agentic tasks. It can navigate software environments, local directories, and codebases without needing human intervention.

On software engineering benchmarks, the model put up solid numbers:

• DeepSWE: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal Bench 4 (CLI Operations): 28.3%

These scores put ML4 ahead of comparable systems like DeepSeek V4 Pro and Qwen3.8 Max. In double-blind human evaluations of code quality by Surge AI, the model scored 3.74 out of 5, ranking behind Anthropic's Claude Opus 5, which led with 4.22 out of 5. Still, ML4 stands out as the strongest open-weights model available, even if it does not quite reach the very top of frontier models.

For automated admin work, ML4 connects with tools like email, spreadsheets, and chat apps. It achieved a 59.9% task completion rate on AutomationBench, and scored 1,393 Elo on AA-Briefcase, which tests high-context processing across documents and spreadsheets.

Local Security Deployment and EU Data Sovereignty

A major challenge with proprietary, cloud-hosted security tools is strict API filtering. For example, after a hack, Hugging Face ran into safety filters on American models while responding to the incident. Because of this, they had to switch to running a Chinese model, GLM 5.2, on their own servers. This shows exactly why a European alternative matters.

Because ML4 offers downloadable weights, security teams can deploy the model locally, keeping sensitive network data within private boundaries. Mistral notes that closed models often refuse to perform legitimate defensive tasks, such as finding and patching software bugs, due to overly strict safety filters. While ML4 actually has a higher refusal rate for malicious cyber requests than any other open model, it is designed to support these defensive security tasks. For safety testing, a version with less moderation is only open to a select group of security partners and governments.

An example of a defense and mitigation workflow looks like this:

[System Intrusion Event]
           │
           ▼
[Telemetry Exported to Private ML4 Instance]
           │
           ▼
[Log Analysis & Defensive Security Assessment]
           │
           ▼
[Automated Patch Generation and Mitigation Code]

ML4 successfully reproduces and patches specific software vulnerabilities 82% of the time. In safety tests, it resists 93.3% of indirect prompt injection attacks on Lakera's b3 benchmark. In capability tests, it successfully solves 93% of Cybench security challenges from CTF competitions.

Additionally, as an EU-based company, Mistral announced they will offer a dedicated European deployment that they run end-to-end. This service operates fully under European law and is independent of other digital service providers. This setup makes it a viable option for highly regulated sectors like banking and public administration, allowing organizations to run ML4 within a secure European framework.

The Reality of "Sovereign" Open Weights

While having open weights sounds like total independence, the reality is more complicated for most organizations.

Even though only 52 billion parameters are active at any one time, running the model still requires loading all 1 trillion parameters into memory. This demands a massive amount of hardware (roughly an estimated terabyte of expensive GPU memory). For a typical setup, that means purchasing and running a heavy GPU server (and likely a second one for backup), plus hiring skilled experts to manage it. For example, the bank HSBC signed a multi-year deal to host Mistral models itself, showing that self-hosting is a true option for the financial sector.

Because of this, hosting ML4 locally is out of reach for average mid-sized businesses or local governments. For them, "sovereignty" will likely mean using Mistral’s European cloud instead of an American provider. This keeps data under European law, but it still means relying on a single vendor.

However, open weights still provide a real benefit: an exit option. Companies can start by using Mistral's API, knowing they can download the model and host it themselves if prices, terms, or international politics change, even if they never use it. This flexibility is a major selling point for risk assessments and public tenders.

In the long run, true independence for the broader market will likely come from smaller, specialized models that we expect Mistral to build on top of ML4. These smaller models can run on a single server or even a laptop, making them far more accessible. In this light, ML4 serves as the powerful foundation that these smaller tools will be distilled from, and the calling card that proves Europe can do it.