Mistral Large 4 is Mistral AI’s new flagship frontier model, released in public preview on October 6, 2026. Nicknamed “Le Chonk”, the model combines a roughly 1 trillion-parameter Mixture-of-Experts architecture with multimodal input, a 1 million-token context window and agentic capabilities aimed at coding, cybersecurity, finance, law and enterprise workflows.
Mistral’s launch post rounds the model to 1 trillion parameters, while its versioned model card lists 1.05 trillion total parameters and 49 billion active parameters, plus a 1.6 billion-parameter vision encoder. The model is available today through Mistral’s public preview API, but the downloadable weights have not been released yet. Mistral says the weights will arrive by the end of October.
The release matters for more than raw model size. Mistral is positioning Large 4 as a European, open-weight alternative to the strongest U.S. and Chinese AI systems, with the ability to run in customer-controlled environments once the weights are released.
Quick take: Mistral Large 4 is a natively multimodal MoE model with 49B active parameters, about 1.05T total parameters, 1M context, 160+ language support and launch API pricing of $1.36 per million input tokens and $4.18 per million output tokens. Mistral is offering a 50% launch discount for two weeks, bringing those rates to $0.68 input and $2.09 output per million tokens during the promotional period.
What is Mistral Large 4?
Mistral Large 4, also referred to as ML4 or Le Chonk, is Mistral AI’s largest and most capable general-purpose model to date.
It is designed to unify several capabilities that previously often required separate specialist models:
- general instruction following;
- reasoning;
- coding;
- agentic workflows;
- multimodal understanding;
- cybersecurity analysis;
- financial and legal knowledge work;
- document and visual reasoning.
Mistral describes ML4 as an open-weight hybrid instruct-and-reasoning Mixture-of-Experts model. The company has not yet published the full architecture details or weights, but says both are coming later this month.
Mistral Large 4 specifications
| Specification | Mistral Large 4 |
|---|---|
| Total parameters | 1.05 trillion according to the versioned model card; launch article rounds to 1T |
| Active parameters | 49 billion |
| Architecture | Granular Mixture-of-Experts |
| Vision encoder | 1.6 billion parameters |
| Context window | 1 million tokens |
| Modalities | Text + image input |
| Languages | 160+ languages |
| Current status | Public Preview |
| Weights | Not yet public; due by end of October |
| API model name | mistral-large-4 |
The 49B active-parameter figure is important because Mixture-of-Experts models do not use all parameters for every token. Only a subset of the model is activated at a time, which can make a very large model more computationally efficient than a dense model with the same total parameter count.
Why 1.05 trillion total parameters does not mean 1.05 trillion active parameters
Large language model parameter counts can be misleading when comparing dense and MoE architectures.
A dense 1T model would involve its full network for each forward pass. Mistral Large 4 instead uses a granular Mixture-of-Experts architecture. According to Mistral’s model card, only about 49 billion parameters are active for a given token.
This means the headline “1 trillion parameters” describes the full capacity distributed across experts, while the active compute per token is much smaller.
The practical benefits can include:
- more specialized internal experts;
- higher total model capacity;
- lower compute requirements than an equally sized dense model;
- better scaling across different tasks;
- potentially more efficient self-hosting once optimized deployments are available.
Actual inference requirements will depend on weight precision, expert routing, batching, hardware and Mistral’s final deployment guidance after the weights are released.
What does the 1M context window mean?
Mistral Large 4 supports a 1 million-token context window.
That gives the model room to process very large collections of text, code, documents and multimodal inputs in one working context.
Potential use cases include:
- large code repositories;
- long legal case files;
- financial filings and research reports;
- multi-document investigations;
- large technical manuals;
- enterprise knowledge bases;
- long-running agent sessions.
A large context window does not guarantee that a model will use every token equally well. Retrieval quality, long-context reasoning and attention over distant information remain separate capabilities that need to be tested in real workloads.
Mistral Large 4 is natively multimodal
ML4 accepts multimodal input, including images.
Mistral lists a 1.6B vision encoder in the model card and highlights visual grounding as one of the model’s strongest areas.
The company says ML4 can reason over images in addition to text and documents, making it useful for tasks such as:
- technical diagrams;
- charts and dashboards;
- screenshots;
- financial documents;
- visual inspection;
- CAD-related analysis;
- multimodal agent workflows.
Mistral claims that in some visual-grounding tasks ML4 can outperform frontier closed models. That is a company benchmark claim rather than a universal independent ranking, so results should be interpreted benchmark by benchmark.
Mistral Large 4 pricing
Mistral Large 4 pricing is currently listed at:
| Usage | Standard price | Two-week launch price |
|---|---|---|
| Input tokens | $1.36 per 1M tokens | $0.68 per 1M tokens |
| Cached input | $0.14 per 1M tokens | $0.07 per 1M tokens |
| Output tokens | $4.18 per 1M tokens | $2.09 per 1M tokens |
Mistral’s changelog says the launch pricing is 50% off for two weeks.
That discount makes the public preview unusually aggressive for a frontier-scale model, especially given the 1M context window and multimodal support.
However, the API price is only one part of cost. Agentic tasks can generate many tool calls, long contexts and large outputs, so real application cost depends on how the model is used.
Are the Mistral Large 4 open weights available now?
No.
This is one of the most important distinctions around the launch.
Mistral calls ML4 an open-weight model because it plans to release the weights, but as of October 7 the public release is still an API preview.
Mistral’s English announcement says the weights will be released by the end of October. Localized versions of the announcement give a late-October date, while Reuters reports that public weight availability is planned for late October after the safety-testing period.
Until then, ordinary developers cannot download the final public ML4 weights for self-hosting.
Why Mistral is delaying the weight release
The delay is primarily about red teaming and cybersecurity evaluation.
Mistral says it is using the preview period to work with cybersecurity leaders, vetted partners and public authorities to evaluate the model’s behavior in real-world settings.
Some trusted partners are receiving access to a version with reduced moderation and expanded cyber capabilities, including preview weights, so they can test defensive and offensive security scenarios more deeply.
The public API remains moderated during this period.
This is a notable compromise between immediate open-weight publication and fully closed deployment: Mistral is committing to downloadable weights but adding a short evaluation window first.
Mistral Large 4 cybersecurity performance
Cybersecurity is one of the main areas Mistral emphasizes in the launch.
The company says ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index and leads non-Chinese open-weight models by a wide margin.
Mistral reports several specific results:
- 82% on an Artificial Analysis test involving reproducing and patching real open-source vulnerabilities;
- 93% of challenges solved on Cybench, a set of security-competition tasks;
- 93.3% resistance on Lakera’s public B3 AI Security Benchmark for indirect prompt-injection attacks.
These numbers come from Mistral’s launch materials and named third-party benchmarks. They do not establish that ML4 is universally the safest or strongest cybersecurity model across every real-world environment.
For background on how agentic systems are tested under adversarial conditions, see Vynula’s AI agent red teaming guide.
Why open weights matter for cybersecurity
Mistral argues that cybersecurity is one of the strongest cases for open-weight deployment.
A security organization may need a model during an active incident, potentially under internal policies that differ from a cloud provider’s safety filters.
With downloadable weights, an organization can theoretically:
- run the model on private infrastructure;
- control moderation policies;
- audit behavior locally;
- keep sensitive incident data on-premises;
- continue using the model even if an external service changes availability.
This flexibility is valuable, but it also increases responsibility. Self-hosted models can remove provider-level guardrails, so organizations need their own monitoring, access controls and abuse prevention.
Vynula’s AI agent guardrails guide covers why local control does not remove the need for runtime safety policies.
How strong is Mistral Large 4 for coding?
Mistral says coding is one of ML4’s strongest general capabilities.
The launch materials include DeepSWE and Terminal Bench 4.0 evaluations, along with internal and external comparisons against other frontier models.
The company also says its human evaluators preferred ML4 in some CAD and STEM tasks while finding it competitive with GLM-5.3 in coding and finance.
Coding benchmarks can vary dramatically depending on repository access, tool use, test harnesses and agent setup, so the most useful comparison will come from reproducible developer workloads after the model’s weights and complete methodology are available.
For a broader market view, Vynula’s best AI models for coding guide tracks the leading coding-focused systems.
Mistral Large 4 is designed for agentic workflows
ML4 is not being positioned only as a chatbot model.
Mistral’s API supports agents, conversations, function calling, structured outputs and built-in tools. The company explicitly describes the model as combining instruction, reasoning and agentic behavior.
That matters because enterprise AI is increasingly shifting from “answer a question” toward “complete a process.”
Examples include:
- researching multiple sources;
- editing spreadsheets;
- calling business tools;
- running code;
- triaging security events;
- planning and executing multi-step tasks.
Vynula’s AI agents guide explains the broader architecture behind these systems.
Finance and legal work are major targets
Mistral is strongly targeting enterprise knowledge work.
The company highlights financial research, spreadsheet editing and legal-agent benchmarks in its launch materials.
It says independent evaluator vals.ai tested ML4 on representative financial and legal tasks and found the model competitive with or ahead of several major systems on those specific evaluations.
On Harvey’s Legal Agent benchmark, Mistral says ML4 outperforms other open models.
These benchmark claims should be treated as task-specific rather than proof that ML4 is a better legal or financial system in every environment.
Real deployments still need domain-specific evaluation, factual verification and professional oversight.
160+ languages and European multilingual training
Mistral says ML4 is natively fluent in more than 160 languages.
The company also says a significant share of the model’s training data is multilingual and includes every official language of the European Union.
This is strategically important because multilingual capability is part of Mistral’s differentiation from frontier models built primarily around English-centric markets.
For European governments and enterprises, language coverage can be as important as benchmark performance when evaluating sovereign or locally deployable AI infrastructure.
Trained on 3,800 NVIDIA Grace Blackwell GPUs in Europe
Mistral says Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company’s own European data centers.
TechCrunch reports Mistral executives describing this as significantly less compute than some major Chinese and closed-model competitors.
The important point is not simply GPU count. Mistral is emphasizing ownership and location of the infrastructure.
The company says the public preview is served on the same Mistral-operated infrastructure and that it will offer deployments in multiple regions, including a European option operated end-to-end under European law.
That ties the model directly to Mistral’s broader AI sovereignty strategy.
What does “AI sovereignty” mean here?
AI sovereignty generally refers to an organization or jurisdiction having meaningful control over the infrastructure, models, data and operational policies it depends on.
Mistral’s approach includes:
- European-controlled training infrastructure;
- open weights;
- self-deployment options;
- European-hosted inference;
- multilingual coverage;
- customer-controlled security policies.
This is particularly relevant for governments, regulated industries and security-sensitive enterprises that do not want all critical AI workloads to depend on a U.S. or Chinese cloud provider.
Reuters frames Large 4 as part of Mistral’s attempt to demonstrate that Europe remains competitive in frontier AI rather than becoming only a customer of American and Chinese systems.
How does Mistral Large 4 compare with Gemini 4 Argon?
Mistral Large 4 and Gemini 4 Argon represent two very different frontier-model strategies.
| Dimension | Mistral Large 4 | Gemini 4 Argon |
|---|---|---|
| Deployment strategy | API preview now, downloadable weights planned | Controlled Google rollout and hosted access |
| Context | 1M tokens | Large-context frontier model; access depends on Google rollout |
| Open weights | Planned by end of October | No public open-weight release announced |
| Primary positioning | Open-weight enterprise, cyber, finance, agentic workloads | Google frontier model for long-horizon professional work |
| Infrastructure story | European-owned training and deployment emphasis | Google’s global AI infrastructure |
This is not a claim that one model is universally better. The key difference is control: Mistral intends to let customers run ML4 themselves, while Google’s frontier strategy remains more tightly hosted.
Mistral Large 4 vs Claude Opus 5.5
Claude Opus 5.5 is a closed Anthropic model focused on high-end reasoning, coding and professional work.
Mistral Large 4 competes for many of the same enterprise use cases, but with a different deployment promise.
Claude emphasizes managed model quality, safety controls and Anthropic’s hosted ecosystem. Mistral emphasizes open weights, self-deployment and customer control.
For a company choosing between them, the decision may depend as much on infrastructure and governance as benchmark scores.
Mistral Large 4 vs GPT-6
OpenAI’s GPT-6 family represents another hosted frontier-model approach.
Vynula’s GPT-6 Sol vs Luna guide covers OpenAI’s current model segmentation.
Mistral’s competitive argument is not simply “our benchmark is higher.” The company is offering a different operating model: customers should eventually be able to own and run the weights themselves.
That makes ML4 especially interesting to enterprises that prioritize:
- on-premises deployment;
- data residency;
- custom safety policy;
- model inspection;
- provider independence.
Why Mistral Large 4 could matter even if closed models remain stronger overall
An open-weight model does not need to win every benchmark to become strategically important.
Enterprise buyers often care about dimensions that standard leaderboards do not capture:
- Can the model run inside our security boundary?
- Can we fine-tune or adapt it?
- Can we control the moderation layer?
- Can we guarantee data residency?
- Can we keep using it if a vendor changes product policy?
- Can we audit the model and deployment stack?
Mistral Large 4 is designed to make those questions part of the frontier-model competition.
Mistral’s safety strategy is unusually relevant because the model will be downloadable
Safety has a different meaning for open-weight models.
A hosted provider can block requests at the service layer, change moderation rules centrally and disable access for abusive users.
Once weights are public, the original provider cannot fully control how every copy is used.
That is why Mistral’s pre-release red-team period is significant.
The company says ML4 performs strongly on indirect prompt-injection defenses and has a higher cyber refusal rate than several other open models in its evaluations.
Those safeguards may still be altered after self-hosting, so organizations deploying the weights will need their own controls.
What Mistral has not published yet
Several important pieces are still missing during the public preview.
- The final downloadable weights.
- Complete architecture details.
- Full post-training methodology.
- Additional benchmark results promised with the weight release.
- Detailed hardware requirements for practical self-hosting.
- Final licensing details associated with the public weight package.
- Independent large-scale evaluations across all the enterprise claims.
Mistral says these details will arrive as the weight release approaches.
Is Mistral Large 4 really open source?
The safer term today is open-weight, not fully open source.
Open-weight means the trained model parameters are intended to be downloadable.
Full open-source AI can imply a broader package including training code, datasets, full methodology and reproducible pipelines. Mistral has not promised that every part of the ML4 training stack will be released publicly.
And as of October 7, even the weights are not yet public.
So the most precise description is: Mistral Large 4 is an open-weight model in public API preview, with the public weights scheduled for later this month.
What developers can use today
Developers can already access the model through Mistral Studio and the public API using mistral-large-4.
The model supports:
- chat completions;
- structured outputs;
- function calling;
- document Q&A;
- batching;
- agents and conversations;
- built-in tools.
That means developers do not need to wait for the open weights to begin testing applications.
The weight release mainly changes the deployment model by enabling direct self-hosting and deeper customization.
Vynula analysis: Mistral is competing on control, not only intelligence
The most important part of Mistral Large 4 is not the 1.05T headline parameter count.
It is Mistral’s attempt to make control a frontier-model feature.
OpenAI, Google and Anthropic largely ask customers to trust a hosted service. They provide powerful models, but the provider owns the core weights and controls the infrastructure.
Mistral is arguing that a frontier-class model should also be something an enterprise can eventually download, deploy, inspect and operate under its own policies.
That changes the competitive framework.
The question is no longer only:
“Which model scores highest?”
It also becomes:
“Which model can I actually own operationally?”
For many consumer use cases, that distinction may not matter. For governments, banks, defense organizations, security teams, manufacturers and critical infrastructure operators, it can matter enormously.
Mistral also has a geographic argument. By training and serving ML4 through infrastructure it operates in Europe, the company is presenting itself as a third frontier-AI path between U.S. hyperscalers and Chinese open-model labs.
That does not mean Europe has suddenly caught or surpassed every U.S. and Chinese model. Mistral’s strongest claims still need more independent evaluation.
But ML4 makes the European AI sovereignty story much more concrete: the region now has a trillion-parameter-class, multimodal, agentic model that is intended to become downloadable rather than permanently locked behind one provider’s API.
If the final weights perform close to the preview API and can be deployed at practical cost, Mistral Large 4 could become one of the most important enterprise open-weight releases of 2026.
What to watch next
- Weight release: whether Mistral publishes the weights on schedule later in October.
- License: what commercial and redistribution terms accompany the final release.
- Self-hosting requirements: how much hardware is needed for useful inference.
- Independent benchmarks: whether third parties reproduce Mistral’s cyber, finance and coding results.
- Safety findings: what the preview red-team period discovers before public weight release.
- Quantized versions: whether Mistral or the community produces practical lower-precision deployments.
- Enterprise adoption: whether regulated industries choose ML4 over hosted frontier models.
- Specialized descendants: Mistral says ML4 will become the base for a new family of optimized models.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 is Mistral AI’s flagship multimodal Mixture-of-Experts model released in public preview on October 6, 2026. It combines reasoning, coding, agentic and visual capabilities.
How many parameters does Mistral Large 4 have?
Mistral’s versioned model card lists 1.05 trillion total parameters and 49 billion active parameters. The launch article rounds the total to 1 trillion.
What is Le Chonk?
Le Chonk is Mistral’s nickname for Mistral Large 4, referring to the model’s very large parameter count.
What is the Mistral Large 4 context window?
The model supports a 1 million-token context window.
How much does Mistral Large 4 cost?
Standard API pricing is listed at $1.36 per million input tokens and $4.18 per million output tokens. Mistral is offering a 50% launch discount for two weeks, reducing those prices to $0.68 and $2.09 respectively.
Are Mistral Large 4 weights available now?
No. The public API preview is live, but Mistral says the public weights will be released by the end of October.
Can Mistral Large 4 understand images?
Yes. It is a natively multimodal model and Mistral lists a 1.6B vision encoder.
How many languages does Mistral Large 4 support?
Mistral says the model is natively fluent in more than 160 languages and was trained on substantial multilingual data.
Was Mistral Large 4 trained in Europe?
Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.
Is Mistral Large 4 fully open source?
The more precise term is open-weight. Mistral plans to publish the weights, but it has not said that every part of the training data, code and pipeline will be released as fully open source.
Primary and independent sources
- Mistral AI — Introducing Mistral Large 4
- Mistral Docs — Mistral Large 4 model card
- Mistral Docs — October 6 model release and launch pricing
- Reuters — France’s Mistral launches AI model it says outperforms some Chinese rivals
- TechCrunch — Mistral’s new 1T model aims to leapfrog closed and open rivals
Reporting note: Benchmark claims are attributed to Mistral unless explicitly identified as independent. The public API is available now, but the public model weights had not been released as of October 7, 2026. Vynula Analysis represents editorial interpretation. Images are illustrative Unsplash photography, not official Mistral product artwork.
