Mistral Large 4: A Trillion Parameters You Can Already Test by API, With Open Weights Due This Month
Mistral Large 4, the model the French company nicknamed Le Chonk, is now in public preview: 1.05 trillion parameters, around 49 billion active per token and natively multimodal. You can use it through the API since October 6, and open weights are promised by the end of the month.
What Mistral Large 4 Is and Why Its Size Matters
Mistral AI opened the public preview of Mistral Large 4 on October 6, 2026. It is the company's largest model to date and it arrived with the nickname Le Chonk, which spread through social media before it reached the technical press. The numbers: roughly 1.05 trillion total parameters, about 49 billion active for every token it generates, a mixture-of-experts (MoE) architecture and native multimodal capability, meaning it handles both text and images in a single model.
A Trillion Total, Fifty Billion Active: What a MoE Actually Is
A mixture-of-experts model splits its network into many specialised parts and activates only a few of them at each step. That is why two figures exist and why they are not interchangeable: the total (1.05 trillion) tells you how much memory the loaded model needs, while the active count (around 49 billion) tells you how much computation each token really costs, which is what drives price and latency. It is also the reason a trillion-parameter model can answer at a price close to a much smaller one.
Multimodal, Long Context and European Training
The context window reaches one million tokens and the input accepts text and images. According to Mistral, training ran in its own European data centres on roughly 3,800 Nvidia Grace Blackwell GPUs, with declared support for more than 160 languages. These are the company's own figures: treat them as a vendor claim, not an independent measurement.
Read also
What Independent Benchmarks Say (and What They Don't)
This is where a launch separates from its press release. The first numbers that matter come from evaluators who do not work for Mistral.
The Best Model Outside the US and China — With Caveats
Artificial Analysis, the firm behind the industry's most cited intelligence index, scored Mistral Large 4 at 38 and described it as the most capable model to come from outside the United States and China. Tom's Hardware framed it the same way: Europe's best effort, with Chinese open weights still ahead. One detail worth keeping in mind is that the score belongs to the preview version, not to a final release.
Chinese Open Weights Are Still Higher in the Ranking
On the same index, DeepSeek V4.1 Flash, a 552-billion-parameter model, landed slightly above it, and the highest-ranked open-weights model on the leaderboard is MiMo-V2.6-Pro at 46. If the leaderboard were the only criterion, the conclusion would be quick: Mistral does not win. The nuance is what the index measures: an aggregate score blends very different tests and says nothing about your specific case.
Price and Cost per Task: the Half of the Comparison Nobody Reads
The reported list price is 1.36 US dollars per million input tokens and 4.18 per million output tokens, with a 50 percent launch discount for the first two weeks, which brings input to around 0.68 dollars. Artificial Analysis puts cost per task at roughly 0.57 dollars under that discount, more expensive than open alternatives such as GLM-5.3-Flash (around 0.25) or DeepSeek V4.1 Flash (around 0.27). In plain terms: Mistral Large 4 does not compete on being the cheapest; it competes on capability, European control and the promise of open weights.
Open Weights at the End of the Month: What It Means for You
Today: Preview API Only
Right now the only access is through the API, using the model identifier mistral-large-4, available in Mistral Studio and already served by third-party gateways such as Otari, Mozilla AI's gateway. The weights, the files that would let you run the model on your own infrastructure, cannot be downloaded yet.
The Memory Math: Why It Doesn't Fit on a Laptop
A trillion parameters at 16-bit precision takes on the order of two terabytes of weights alone. That is an order-of-magnitude estimate, not an official figure, but it is enough to make the point: a MoE requires the whole model's memory to be reachable, not just the experts that fire on each step. Self-hosting here means a multi-GPU server, not a desktop machine. If you come from the local-model world, this is not your next stop: a small quantised model still makes far more sense for that, as we covered in the guide to running LLMs locally with Ollama.
What Isn't Published Yet: License, Quantizations, Requirements
The promised window is "the end of October," and some outlets point to the 27th, but no day is officially confirmed. Nor has the license been published, nor whether official quantised versions will exist, nor the hardware requirements. In other words, the open part of this story has not been written yet, and anyone promising the opposite today is speculating.
How to Test It This Week, Step by Step
Get Access and Make Your First Call
If you already have a Mistral Studio account, the call is the usual chat completions request with a different model identifier:
curl https://api.mistral.ai/v1/chat/completions \
-H "Authorization: Bearer $MISTRAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"mistral-large-4","messages":[{"role":"user","content":"Summarise this text in two sentences."}]}'
Before you wire it into production, confirm the identifier and the endpoint in the official documentation: preview aliases change, and the nickname is not a valid identifier.
A Five-Prompt Test of Your Own (Better Than the Leaderboard)
A general ranking does not translate directly into your product. Five tests with your own data say much more:
- A coding task with a real bug from your repository and a known correct answer.
- A multimodal question about a specific image, checking an exact figure that appears in it.
- A long-context task, to see whether it keeps the thread or starts filling gaps.
- An agent task that must return structured JSON, validating the output against a schema.
- The same task on a smaller open model, to compare cost per completed answer rather than raw quality.
What to Measure: Correctness, Format, Latency and Cost per Task
Measure four things: whether it gets the answer right, whether it respects the format you need, how fast it feels, and what each completed task costs. That last number is the one that decides in production, and it comes from multiplying your real token usage by the published price, not from reading the top of a table.
What This Launch Says About the Market
Europe, Open Weights and Dependence on US Providers
Mistral's argument is not only technical: it is that a model this size was trained entirely in Europe and that its weights will be released. For companies with strict rules about where their data is processed, that weighs more than a couple of index points. Without overstating it: one company's launch does not move the market on its own, and Chinese open weights still set the pace on capability per dollar.
What to Watch Over the Next Few Weeks
Four things: the exact weight release date, the license, whether official quantizations appear, and what independent evaluators say once they test the final version rather than the preview. And one practical note for anyone planning to self-host: moving enormous weights from storage to the GPU is not a networking problem, it is a model-loading problem. That is the subject of the article on reducing start latency in LLM inference with NVIDIA Run:ai Model Streamer.
Conclusion
What to do today: if the model interests you, test it through the API with your own five-prompt set while the launch discount is live, and record the cost per completed task. What not to do yet: buy GPUs on hype, because the weights do not exist as a download and the license is unpublished. When they land, we will cover part two here: what you can actually run, with how much memory and at what price. If you follow the open-model ecosystem, this blog is a good place to keep reading.
