Mistral Large 4 Explained: What a 1T Open-Weight Model Changes

By Jaspal Singh October 7, 2026 Updated: October 7, 2026

Mistral Large 4 is a 1-trillion-parameter, open-weight AI model from France's Mistral AI that activates only about 49 billion parameters per token. It is in paid API preview now, with downloadable weights promised by the end of October 2026. For businesses, it is the strongest model they may soon be able to run on their own hardware outside China.

What Happened

On October 6, 2026, Mistral AI launched a public preview of Mistral Large 4, which the company calls "ML4" and, unofficially, "le Chonk". According to Mistral's launch announcement, the model has 1 trillion total parameters and 49 billion active parameters, is natively multimodal (it reads text and images), and is fluent in more than 160 languages, including every official language of the European Union. Mistral's model documentation gives slightly more precise figures: 1.05 trillion total parameters, 52 billion active, a 1.6-billion-parameter vision encoder and a context window of one million tokens.

The model was trained on about 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centres. The preview is available through Mistral Studio under the API name mistral-large-4, priced in the documentation at $0.68 per million input tokens, $0.07 per million cached input tokens and $2.09 per million output tokens. Mistral says it "will release the weights by the end of the month". Until then, it is red-teaming the model with cybersecurity firms, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities.

The launch follows Mistral's €3 billion Series D round, which the company describes as the largest equity raise by a European technology company. Coverage from CNBC, Wired, The Register, ZDNET, The Decoder and others centred on one claim: that ML4 is the most capable open-weight model built outside China.

Why It Matters

Most of the world's strongest AI models are closed. You can rent them through an API, but you cannot download them, inspect them or run them on your own servers. The strongest open-weight models of the past year have mostly come from Chinese labs such as DeepSeek, Alibaba's Qwen team, Moonshot and Z.ai. That has left European and American organisations that need to keep data on their own hardware choosing between weaker Western open models and Chinese ones, and some of them cannot use Chinese models for procurement or compliance reasons.

ML4 is Mistral's attempt to fill that gap. If the weights arrive as promised, a bank, hospital, defence contractor or government ministry could run a near-frontier model entirely inside its own data centre, fine-tune it on private documents, and never send a prompt to a third party. Mistral also stresses that its own hosted version is operated in Europe under European law, independently of other digital service providers, which is a direct pitch to buyers worried about US cloud jurisdiction.

The second reason it matters is cybersecurity. Mistral says ML4 ranks in the top five models worldwide on the Artificial Analysis Cyber Index and scores 82% on a vulnerability reproduction test, a task where it says closed models from Anthropic and OpenAI score close to zero because their safety policies make them refuse. Mistral's argument is that "defending software often starts with proving that a flaw is real", and that refusal-heavy closed models get in the way of legitimate security teams. Critics will point out that the same capability helps attackers, and that once weights are public no one can take them back.

It is worth keeping the scale of the claim in proportion. The Decoder reports that ML4 scores 38 on the Artificial Analysis Intelligence Index, against 58 for Claude Opus 5.5. In a blind human coding evaluation run by Surge AI and cited by Mistral, ML4 came second of five with 3.74 out of 5, behind Claude Opus 5 at 4.22. ML4 is a strong open model, not a frontier leader.

How It Works

Mixture of experts: why 1 trillion parameters does not mean 1 trillion calculations

ML4 uses a mixture-of-experts (MoE) design. Instead of one giant neural network where every parameter takes part in every prediction, the model contains many smaller sub-networks, called experts, plus a small routing network. For each token the model processes, the router picks a handful of experts and only those run. Mistral calls its design "granular" MoE, meaning it uses many small experts rather than a few large ones, which lets the router combine specialists more finely.

The practical effect is that ML4 stores the knowledge of a 1-trillion-parameter model but spends roughly the compute of a 49-to-52-billion-parameter model on each token. That is why Mistral can charge less than a dollar per million input tokens for a model this large. Mistral's previous flagship, Mistral Large 3, used the same idea at a smaller scale: 675 billion total parameters with 41 billion active.

The catch: memory

MoE saves compute, not memory. Every expert has to be loaded and ready, because the router might pick any of them for the next token. A rough calculation shows the problem: 1.05 trillion parameters stored at 8 bits each is about 1 terabyte of weights before you add the memory needed for a long context. Even squeezed to 4 bits, it is around half a terabyte.

That rules out laptops, gaming PCs and single-GPU workstations. The Register reports that the model is designed to run on eight-GPU servers such as Nvidia's HGX B300 or AMD's MI355X systems, the kind of hardware that costs several hundred thousand dollars. For most people, "open weights" will mean using ML4 through a cloud host or Mistral's own API, while self-hosting will be realistic for large companies, research labs, governments and specialist hosting providers.

Reasoning, vision and a one-million-token context

ML4 is a hybrid instruct-and-reasoning model: the same model can answer quickly or work through a problem step by step. Its vision encoder lets it read charts, scanned documents, technical drawings and satellite images, and Mistral claims it beats OpenAI's GPT-6 Astra on one visual grounding benchmark (42% versus 41% on Dense 200). The one-million-token context window means it can take in several long contracts or a large codebase in one request, although the cost and memory of filling that window are substantial.

Post-training with reinforcement learning

Mistral describes a post-training pipeline that combines supervised fine-tuning with large-scale reinforcement learning across composable environments for chat, science, tool use, factuality and safety. It says the system generates around 33 billion tokens of rollouts a day and that the reinforcement learning run is "still in flight", which is why the company expects the model to keep improving and has hinted at follow-up versions.

What's Still Unknown

The licence. Mistral Large 3 shipped under the permissive Apache 2.0 licence. The ML4 announcement and documentation we read do not say which licence the weights will use. That single detail decides whether companies can use, modify and resell the model freely, or whether there will be restrictions on commercial or security use.

Whether the public weights match the preview. Partners currently get a version with reduced moderation for cyber work. It is not yet clear whether the downloadable weights will carry the same capabilities, extra safeguards, or a separate restricted release.

Independent benchmarks. Most headline numbers come from Mistral itself. The Decoder notes that comparisons with Chinese models should be taken with a grain of salt until third parties repeat them on the final weights.

Final pricing and specifications. Mistral's blog and documentation differ slightly on active parameters (49 billion versus 52 billion), and preview pricing may change at general availability. The exact release date of the weights is also only "end of the month".

Frequently Asked Questions

What is Mistral Large 4?

Mistral Large 4 is the newest flagship AI model from Mistral AI, a Paris-based company. It is a mixture-of-experts model with about 1 trillion total parameters and roughly 49 to 52 billion active per token. It handles text and images, supports more than 160 languages and a one-million-token context, and is in public API preview from October 6, 2026.

Why is Mistral Large 4 called Le Chonk?

"Le Chonk" is an unofficial nickname Mistral gave the model in its own launch post, playing on internet slang for something big and chunky. It refers to the model's size, at roughly one trillion parameters the largest Mistral has built. The official name used in the API and documentation is Mistral Large 4, shortened to ML4.

Is Mistral Large 4 open source?

It is described as open-weight, which means Mistral plans to publish the trained model files so anyone can download and run them. That is not the same as fully open source, because training data and code are not included. The weights are due by the end of October 2026, and Mistral has not yet said which licence they will use.

Can I run Mistral Large 4 on my own computer?

Not on a normal PC or laptop. Even though only about 49 billion parameters are active per token, all 1 trillion must sit in memory, which means roughly half a terabyte to a terabyte of fast GPU memory. Reports point to eight-GPU servers such as Nvidia HGX B300 systems. Most users will reach it through an API or a hosting provider.

How much does the Mistral Large 4 API cost?

Mistral's documentation lists preview pricing of $0.68 per million input tokens, $0.07 per million cached input tokens and $2.09 per million output tokens. That is far cheaper than most frontier closed models, which reflects the mixture-of-experts design. Prices are labelled preview pricing, so they may change when the model reaches general availability or after the weights are released.

Is Mistral Large 4 better than GPT-6 or Claude?

Generally no. Independent scoring reported by The Decoder puts it at 38 on the Artificial Analysis Intelligence Index, against 58 for Claude Opus 5.5, and a blind coding test cited by Mistral ranked it behind Claude Opus 5. It does beat leading closed models on a few specific tasks, such as visual grounding and security work that closed models refuse.

Why does Mistral emphasise cybersecurity?

Mistral argues that defenders need models willing to reproduce real vulnerabilities, analyse malware and write detection rules, tasks that safety filters in closed models often block. ML4 scores 82% on a vulnerability reproduction test and ranks in the top five worldwide on a cyber index. The same capabilities could help attackers, which is why Mistral is red-teaming before releasing the weights.

Who should care about Mistral Large 4?

Organisations that cannot send data to US or Chinese AI providers have the most to gain: European governments, banks, healthcare groups, defence suppliers and security firms. Developers also benefit from a cheap, capable API and a future model they can fine-tune. Everyday chatbot users will notice less directly, though Mistral's own chat assistant offers access to it.