Mistral AI is a French company that opened a public preview of its new large model, Mistral Large 4, on October 6, 2026. The model has approximately 1 trillion parameters and a 1 million-token context window. For now, it can only be used through Mistral Studio and the API, but the company promises to release open weights by the end of October.
What Is Mistral AI
Mistral AI is an artificial intelligence company headquartered in Paris and is often described as the main European alternative to American labs such as OpenAI. The company develops its own language models, releases some of them with open weights, and sells access to them through the API and Mistral Studio.
On September 8, 2026, Mistral AI announced a €3 billion Series D funding round led by Samsung Electronics. The company’s valuation exceeded €21 billion — almost twice the €11.7 billion valuation it had in 2025. Other investors mentioned include Scaleup Europe Fund (EQT) and PSG Equity, while Microsoft and Nvidia remain existing investors. According to the company, the funds will be used for data centers and computing capacity. According to Euronews, Mistral AI technologies are used by more than 125 enterprises in 20 countries.
People often ask whether the company is connected to OpenAI: no, they are independent competitors. Mistral AI is not owned by OpenAI and is not part of it.
What Is Mistral Large 4
According to Mistral’s documentation, the model has 1.05 trillion parameters in total, with about 52 billion active for each token (some early reports mentioned 49 billion). It uses a mixture-of-experts architecture: only part of the model is activated for each piece of text, which reduces compute requirements.
The model is multimodal — it accepts text and images — and, according to the company, supports more than 160 languages. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers. At the same time, the company notes that the model is still undergoing red-teaming (testing for misuse) and is “improving rapidly.”
What Is Confirmed, What Is Promised, and What Is Still Only Rumor
| Status | Fact |
|---|---|
| Confirmed by Mistral | Preview from October 6; access through Mistral Studio and API; 1 million-token context |
| Confirmed | No consumer app and no weight download |
| Promised | Open weights “by the end of the month” (October 27 date comes from The New Stack, not Mistral) |
| Not published | License: reports say it will be custom rather than Apache 2.0, as with Large 3 |
| Not confirmed | Claims that the model “outperforms Claude, ChatGPT and Gemini” |
How Much Mistral Large 4 Costs
During the preview, discounted pricing applies: $0.68 per million input tokens and $2.09 per million output tokens — half the regular prices of $1.36 and $4.18. After the preview ends, the standard rates will apply, so use those figures when planning a budget.
Mistral Large 4 Benchmarks: What the Numbers Say
Almost all results are self-reported. On the DeepSWE test (programming), Mistral reports 61.7%, while the live leaderboard shows Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash at around 74%. In a blind coding test conducted by Mistral itself, Large 4 scored 3.74 out of 5 and placed second out of five, behind Claude Opus 5 (4.22).
The conclusion: the model is clearly not “the best in the world” at coding, but it looks strong for an open solution, especially considering the price and the possibility of self-hosting if the weights are indeed released.
Can You Run Mistral Large 4 at Home
The short answer is no. Even with 4-bit quantization, the full model takes up approximately 500 GB, which is more than a single RTX 5090 graphics card can hold. So self-hosting is currently a task for data centers and companies, not home enthusiasts. For regular users, smaller models that Mistral and other companies release openly may be more interesting.
Who Mistral Large 4 Is For
- Companies with data residency requirements looking for a European solution
- Developers who want a low-cost API with long context
- People following open models and waiting for the weights and license
For most chatbot users, the conclusion is simple: there is no need to switch tools yet — it is better to wait for independent tests and the weights release.
In Brief: The Key Points
- October 6, 2026 — public preview of Mistral Large 4 (Mistral Studio and API)
- ~1.05T parameters, ~52B active, 1 million-token context
- Multimodal (text + images), more than 160 languages
- Preview pricing: $0.68 / $2.09 per million tokens
- Weights promised by the end of October, license not yet published
- Benchmarks are self-reported; it cannot run on a home PC (~500 GB)
FAQ
What is Mistral AI? A French company that develops language models and sells access to them through an API; some models are released with open weights.
What is Mistral Large 4? A large multimodal model from Mistral AI with ~1 trillion parameters and a 1 million-token context window, available in preview through Mistral Studio and the API.
When was Mistral Large 4 released? The public preview opened on October 6, 2026.
Can Mistral Large 4 be downloaded? Not yet. Mistral promises to release open weights by the end of October, but the license has not yet been published.
Is Mistral Large 4 better than ChatGPT, Claude and Gemini? This has not been confirmed. According to self-reported data, the model trails the leaders on DeepSWE by about 12 points.
Will Mistral Large 4 run on a home computer? No. Even in 4-bit form, it takes up around 500 GB.
