Mistral Large 4: What the Preview Gives You
Mistral Large 4 is in public preview: 1T parameters, 49B active, $1.36 per million in. What the preview gives you now and what waits for weights.
Mistral released Mistral Large 4 into public preview on 6 October 2026. It is a 1 trillion parameter mixture-of-experts model with 49 billion parameters active per token, natively multimodal, available through the API on Mistral Studio today, with open weights promised by the end of October. Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens.
That last sentence is the one worth sitting with. A model you can download in three weeks is a different purchasing decision from a model you can only rent, and Mistral has put both on the table at once.
What is actually in the release
The headline numbers from Mistral's announcement:
1.05 trillion total parameters, 49 billion active. A granular mixture-of-experts design, so you pay inference cost closer to a 49B model than a 1T one. If that split is new to you, what is mixture of experts covers the mechanism.
Hybrid instruct and reasoning. One model that handles both ordinary instruction following and longer reasoning, rather than a separate reasoning variant.
Natively multimodal. Vision is part of the model rather than a separate pipeline bolted on at the edge.
Trained in Europe. Mistral says it trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters.
Nicknamed le Chonk. Officially. This is not a joke someone else made.
The benchmark numbers, with the price attached
Mistral published scores across coding, terminal use, security and vision. Read them next to the price rather than on their own:
Benchmark | Mistral Large 4 | What it measures |
|---|---|---|
DeepSWE v1.1 | 61.7% | Resolving real software issues |
Terminal Bench 4.0 | 28.3% | Multi-step work in a shell |
Cybench | 93% | Reproducing known vulnerabilities |
AutomationBench | 59.9% | End-to-end task automation |
Dense 200 | 42% | Visual grounding, against 41% for GPT-6 Astra |
The vision result is the narrowest margin and the one Mistral chose to put a competitor next to. The Terminal Bench figure is the one to take seriously in the other direction: 28.3% means roughly seven in ten multi-step terminal tasks still fail, which is normal for the category and a useful corrective to the trillion-parameter headline. Benchmark tables deserve the scepticism in how to tell when an AI benchmark score is misleading before they change anyone's stack.
What the preview label means for you
A preview API is a model you can test but should not pin a product to. Mistral is explicit that during the preview it is red-teaming the model with cybersecurity partners and state authorities, who get access to a version with reduced moderation and expanded cyber capabilities. That is a sensible disclosure, and it also tells you the production behaviour is still moving.
Three practical consequences:
Treat preview output as a signal about capability, not as a contract about behaviour. Prompts tuned against a preview often need re-tuning at general availability.
Do not build a cost model on preview pricing. Preview prices are a marketing position as much as an economic one.
If you do start testing now, keep the evaluation set you build. It is the thing that transfers when the weights land.
The open weights question
Mistral says weights follow by the end of this month. If that holds, the decision for anyone weighing self-hosting changes shape: you get to measure the model on the API first and only then decide whether running it yourself is worth the operational cost. That sequencing is unusual and genuinely useful.
It is not automatically cheaper. A 1T-parameter model, even a sparse one, needs serious hardware to serve, and the arithmetic in is self-hosting an open weight model cheaper than an API tends to favour the API until volume is high and steady. What open weights buy reliably is control: no deprecation on someone else's schedule, and data that never leaves your infrastructure. For teams with that constraint, what is a sovereign AI model explains why a European training story is a procurement argument and not just a press line.
Should you switch
Not this month, on this evidence. The honest position on a preview release is that it is a candidate, not a replacement. Add it to whatever comparison you already run, price it against your current provider using the method in how to compare AI API pricing across providers, and wait for the weights before making a structural decision.
What is worth noting is the shape of the release rather than the model. A frontier-scale model with published weights three weeks behind the API, trained and served inside the EU, at under half the output price of the frontier tier, is a different competitive proposition from the one the market had last month. Keeping track of which of these claims survive contact with real workloads is the job described in how to keep up with AI news without reorganising your week around it.
FAQ
When do Mistral Large 4 weights come out?
Mistral says by the end of October 2026. The API preview is available now through Mistral Studio; the weights have not been released at the time of writing.
How much does Mistral Large 4 cost?
Preview pricing is $1.36 per million input tokens and $4.18 per million output tokens. Preview pricing is not a guarantee of general availability pricing.
How many parameters does Mistral Large 4 have?
About 1.05 trillion in total, with 49 billion active per token. The active count is what drives inference cost, not the total.
Is Mistral Large 4 open source?
Not yet, and open weights is not the same as open source. Mistral has committed to releasing the weights; it has not published licence terms at the time of writing, so treat any commercial-use assumption as unconfirmed.
How did this land?
About the author

Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.


