Mistral Large 4: Features, Benchmarks, Pricing and How to Use It (2026 Guide)

Mistral Large 4 is the new flagship model from Paris-based Mistral AI, unveiled on 6 October 2026 as a public preview. Nicknamed “Le Chonk”, it is a mixture-of-experts model with around one trillion parameters, it understands images as well as text, and Mistral has promised to release its weights by the end of October. This guide explains what Mistral Large 4 is, what the benchmarks actually show, how much the API costs, and how to start using it today in Vibe or through Mistral Studio.

Everything below was checked on 7 October 2026 against Mistral’s announcement, its model documentation and coverage from SiliconANGLE, The Next Web and AI News. Because the model is still in preview, specifications and prices may change before the open-weight release.

What Is Mistral Large 4?

Mistral Large 4 (ML4) is the successor to Mistral Large 3 and the most capable model Mistral has released. The key facts from Mistral’s announcement and documentation:

The “mixture-of-experts” design is why such a large model can be practical: only about 49 billion of its parameters are active for each token, so it needs far less computing power per request than a dense model of the same total size.

Mistral Large 4 Benchmarks: What the Numbers Say

Mistral and independent outlets highlighted several results. These are the company’s own figures unless stated otherwise, so treat them as a starting point rather than a verdict.

Cybersecurity

This is where Mistral is most confident. The announcement cites a top-five global position on the Artificial Analysis Cyber Index, 82% on a vulnerability reproduction test and 93% on Cybench exercises. Co-founder Guillaume Lample told The Next Web that the cyber-defence capabilities are meant to help enterprises and governments defend against attackers who jailbreak closed models.

Coding

The coding picture is more mixed. Mistral reports 61.7% on DeepSWE, 28.3% on Terminal-Bench 4 and 49.8% on a Coding Agent Index. The Next Web says that on DeepSWE it narrowly beats Zhipu’s GLM-5.3 (61%) and DeepSeek-V4-Pro (57%). SiliconANGLE notes that it still trails the leading frontier models on coding, and AI News reports a blind human evaluation in which it scored 3.74 out of 5, second of five models, behind Claude Opus 5 at 4.22.

Vision and agentic work

• Visual grounding: 42% on Dense 200 versus 41% for GPT-6 Astra, according to Mistral, and 73% on DIOR-RSVG versus 53% for DeepSeek Flash and 55% for Kimi K3, according to The Next Web.
• Office and finance tasks: 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase, ahead of DeepSeek V4 Pro according to SiliconANGLE; The Next Web reports a tie with DeepSeek-V4-Pro on FinWorkBench at 67%.

Our reading: Mistral Large 4 looks like the strongest open-weight model from a European company and is competitive with the best Chinese open models, especially for security and visual tasks. It does not claim to beat the top closed models across the board. If you want a closer look at one of those, our guide to Claude Sonnet 5.5 covers a leading closed alternative.

Mistral Large 4 Pricing

Mistral’s sources currently show two sets of prices, so check the console before you budget:

The documentation figures are exactly half the announced prices, which suggests a preview discount, but Mistral has not explained the difference on the pages we checked. Either way, Mistral Large 4 is priced well below most closed frontier models. For comparison, our DeepSeek API guide covers one of the cheapest alternatives, and our report on GPT-6 Sol and Luna pricing shows where OpenAI sits.

If you only want to chat with the model, you do not need the API. Mistral says Large 4 is available in Vibe, its consumer assistant (formerly Le Chat), which has a free plan and a Pro plan. Our Mistral Vibe guide explains the plans and how to get started.

How to Use Mistral Large 4

Option 1: In Vibe (no code)

1. Sign in at chat.mistral.ai or open the Mistral Vibe app.
2. Start a task in Work mode and choose the model if a model selector is shown for your plan.
3. Try a task that plays to its strengths: summarise a long document, analyse an image or chart, or review a piece of code for security issues.

Mistral has not published which Vibe plans include Large 4 during the preview, so if you do not see it, check Mistral’s help centre or try again after the wider rollout.

Option 2: Through the Mistral API (developers)

Mistral’s documentation says API access in Mistral Studio is enabled by default with no credit card required, in a free mode with usage and rate limits.

1. Create an API key: go to console.mistral.ai, open API Keys, click Create new key, give it a name and expiry date, and copy it immediately. Mistral warns that the full key is shown only once.
2. Store the key as an environment variable called MISTRAL_API_KEY.
3. Send a request to the chat completions endpoint with the model ID mistral-large-4:

curl https://api.mistral.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -d '{
    "model": "mistral-large-4",
    "messages": [
      {"role": "user", "content": "Summarise the EU AI Act for a small software company in five bullet points."}
    ]
  }'

In Python, install the official SDK with pip install mistralai:

import os
from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

response = client.chat.complete(
    model="mistral-large-4",
    messages=[{"role": "user", "content": "Explain mixture-of-experts in plain English."}],
)
print(response.choices[0].message.content)

Mistral’s model page lists support for structured outputs, function calling, document Q&A, batching, agents and built-in tools, so the same model can power chatbots, document workflows and agents. Document search also needs an embedding model; our embedding model comparison covers Mistral Embed alongside open alternatives. Mistral’s quick-start examples use the alias mistral-large-latest; if you want to stay on Large 4 specifically, use the explicit ID rather than the alias, which may point to a different version over time.

Option 3: Self-host the open weights (from late October)

Once the weights are published, organisations will be able to run Mistral Large 4 on their own servers. Mistral has also said it will publish architecture details and its post-training method at the same time. The licence had not been confirmed on the pages we checked; Mistral Large 3 used Apache 2.0, but do not assume Large 4 will use the same terms until Mistral confirms them.

Be realistic about hardware. As a rough estimate, 1.05 trillion parameters at 4-bit precision would need around 525 GB just for the weights, before any memory for the context. That is data-centre territory, not a gaming PC. If you want a model you can run at home, our guide on how to run an LLM locally covers smaller open models that fit on a laptop or desktop.

Why Mistral Large 4 Matters for UK and EU Users

• European provider: Mistral is a French company and says the model was trained in European data centres. It offers deployment in multiple regions, including an end-to-end European option, which matters for organisations with data-residency requirements under GDPR.
• Language coverage: support for more than 160 languages, including every official EU language, makes it a strong candidate for multilingual customer service and document work in Germany, France, Spain, Italy, Poland, the Netherlands and beyond.
• Open weights: the ability to self-host gives regulated organisations full control over where data is processed.
• Price: per-token prices are well below most closed frontier models.

Strengths and Weaknesses

Strengths

• Open-weight release promised, unlike most frontier models
• Very strong reported cybersecurity results
• Native image understanding and a 1M-token context window
• Broad European language support and EU deployment options
• Competitive API pricing, with a free API tier for testing

Weaknesses

• Still a preview: specifications and prices may change
• Trails the best closed models on coding, according to SiliconANGLE and AI News
• Licence not yet confirmed
• Far too large for consumer hardware once the weights are released
• Conflicting price information between Mistral’s own pages

Who Should Use Mistral Large 4?

Good fit: European businesses that want a capable model from an EU provider; security teams interested in its cyber-defence results; developers building multilingual or document-heavy applications; and organisations that plan to self-host once the weights are available.

Maybe not yet: developers whose main priority is the very best coding performance, and anyone who needs fixed prices and a stable API before the preview ends.

Frequently Asked Questions

When was Mistral Large 4 released?

Mistral unveiled Mistral Large 4 as a public preview on 6 October 2026. The open weights are due by the end of October 2026.

Is Mistral Large 4 open source?

Mistral describes it as an open-weight model, and its model page lists the licence category as open. The weights are due by the end of October, but the exact licence terms had not been confirmed when we checked.

How many parameters does Mistral Large 4 have?

About 1 trillion in total (Mistral’s model page says 1.05 trillion), with 49 billion active per token.

How much does Mistral Large 4 cost?

Mistral’s announcement lists $1.36 per million input tokens and $4.18 per million output tokens. The model documentation page currently shows $0.68 input, $0.07 cached input and $2.09 output. Check the Mistral console for the rate that applies to your account.

Is Mistral Large 4 on Hugging Face?

Not yet. Mistral says it will release the weights by the end of October 2026; at that point, check Mistral’s official Hugging Face organisation.

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

Not overall, according to current reports. It edges out GPT-6 Astra on one visual grounding benchmark and performs strongly on cybersecurity, but SiliconANGLE and AI News report that it trails the leading closed models on coding.

Conclusion

Mistral Large 4 is Europe’s most ambitious open-weight model so far: roughly a trillion parameters, image understanding, a 1M-token context window and standout cybersecurity results, at prices well below most closed rivals. You can try it today in Vibe or through the Mistral API with the model ID mistral-large-4, and self-host it once the weights arrive at the end of October. Treat the benchmarks as Mistral’s own claims, check the final licence and prices when the preview ends, and test it on your own tasks before switching production workloads.

Sources: Mistral – Introducing Mistral Large 4; Mistral Large 4 model documentation; Mistral first API request; Mistral Studio API key guide; SiliconANGLE; The Next Web; AI News.