Mistral Large 4 is the new flagship model from Mistral AI, the French startup that has become Europe's main contender in generative AI. Released as a public preview on 6 October 2026, it is a natively multimodal, 1-trillion-parameter mixture-of-experts model with a one-million-token context window, and Mistral has promised to release its open weights by the end of October. Here is what it is, how it stacks up against GPT, Claude and Gemini, what it costs and how to try it.
What is Mistral Large 4? Key specs
Based on Mistral's announcement and model card, these are the essentials:
- Size: 1.05 trillion total parameters, of which only 52 billion are active per step. The Hugging Face listing gives 49 billion active excluding embedding and output layers, which is why both figures are in circulation.
- Architecture: a mixture-of-experts (MoE) model that merges instruction-following, reasoning and agentic behaviour in a single model.
- Multimodal: understands text and images (via a 1.6-billion-parameter vision encoder) and answers in text. It is pitched at documents, charts and geospatial imagery.
- Context: up to 1 million tokens, enough for long contracts, codebases or reports in one go.
- API features: function calling, structured outputs, document Q&A, batching and agents.
- Training: from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres, with data in more than 160 languages including every official EU language.
For scale, its predecessor Mistral Large 3 had 675 billion total and 41 billion active parameters, according to VentureBeat.
Is Mistral Large 4 open source? Weights and licence
Not yet downloadable. Mistral launched an API preview first and says it will release the weights "by the end of the month". VentureBeat reports a target date of 27 October 2026, and Hugging Face already hosts a placeholder page named Mistral-Large-4-1T-A52B. In the meantime the model is being red-teamed with cybersecurity partners and state authorities.
The licence is less clear. Mistral has not published it, and VentureBeat expects a custom Mistral licence rather than necessarily Apache 2.0. Read it carefully once it lands if you plan commercial use.
A practical warning: open weights do not mean laptop-friendly. One trillion parameters at 8-bit precision is roughly 1 TB of weights alone, which means multi-GPU servers. If you want something you can run yourself, see our guide to open-source Claude alternatives you can install on your own server.
Mistral Large 4 vs GPT, Claude and Gemini
Mistral has published plenty of numbers, but most are self-reported, and the preview is still undergoing reinforcement learning. With that caveat, the most useful data points are:
| Benchmark | Mistral Large 4 | Reference |
|---|---|---|
| DeepSWE v1.1 (coding agents) | 61.7% | GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 around 74% on the public leaderboard (per VentureBeat) |
| Blind human coding eval (Surge AI, 1-5) | 3.74 | Claude Opus 5: 4.22; Kimi K3: 3.59 |
| Vulnerability reproduce-and-patch | 82% | Several closed models near 0% because they refuse the task |
| Dense200 visual grounding | 42% | GPT-6 Astra: 41% |
The honest reading: in coding and general reasoning, Mistral Large 4 does not catch the leading closed models from OpenAI, Anthropic and Google. Its real target is to be the best open-weight model outside China, competing with DeepSeek, Qwen, GLM and Kimi, while standing out in specific niches such as cybersecurity, finance, legal work, manufacturing and image analysis. For a head-to-head of the big assistants, see our Claude vs ChatGPT vs Gemini comparison.
Mistral Large 4 pricing and how to try it
Standard API pricing is $1.36 per million input tokens and $4.18 per million output tokens ($0.14 for cached input). For the first two weeks after launch Mistral ran a 50% discount ($0.68 and $2.09), so check the pricing page for the current rate when you test it.
To get started:
- Sign up for Mistral Studio (console.mistral.ai), the developer platform.
- Try the model in the playground or create an API key.
- Call the chat completions endpoint with the model ID mistral-large-4. Function calling, structured outputs and the Agents API are supported.
What about Mistral's chatbot? Le Chat has been renamed Vibe and has a free plan, but at the time of writing Mistral has not publicly confirmed which model Vibe uses by default or whether Large 4 is available to free users. For a guaranteed Large 4 experience, use Studio or the API.
Why Mistral Large 4 matters for European users and businesses
Mistral's pitch is not only technical. The model's tagline is "Forged in Europe. Built for AI sovereignty", and that has practical consequences:
- Data stays in Europe: the preview runs on Mistral's European infrastructure, the company offers a European deployment it operates end to end, and enterprise customers can get zero-data-retention options.
- GDPR and compliance: working with an EU-based provider under EU law makes conversations with your data protection officer easier, though you still need to review the data processing agreement.
- Full control with open weights: once released, the model can run in a private cloud or on premises, something impossible with GPT, Claude or Gemini. For government, banking and healthcare, that is often decisive.
- Strong European language coverage: multilingual training spans every official EU language, Spanish included.
There is a political backdrop too. Mistral closed a €3 billion Series D in September at a valuation above €21 billion, and French President Emmanuel Macron has called its approach "a third way in AI", as TechCrunch reports. If you build RAG systems over internal documents, an open European model with a one-million-token context window deserves a serious look.
Sources: Mistral Large 4 announcement, model docs and pricing and TechCrunch.
Frequently asked questions
Is Mistral Large 4 free?
Not via the API: it costs $1.36 per million input tokens and $4.18 per million output tokens. Once the open weights are out you will be able to download and run it without per-token fees, but you will need very powerful servers.
How many parameters does Mistral Large 4 have?
About 1.05 trillion in total, with 52 billion active per token (49 billion excluding embeddings and the output layer).
Is Mistral Large 4 better than ChatGPT or Claude?
Not overall. Its own published numbers put it behind the top closed models in coding, but it competes with the best open models and stands out in cybersecurity and visual analysis.
When will Mistral Large 4 weights be released?
Mistral has promised them by the end of October 2026; VentureBeat reports a target date of 27 October.
