Mistral says its new model is open weight and cheap. Priya Chen on what the evidence shows so far.
- Verified fact: Mistral lists mistral-large-4 in public preview as of 6 October 2026 with a 1 million token context window.
- Attributed claim: the headline benchmarks are reported by Mistral, and the weights and licence are not yet published.
- Analysis: treat it as a model to test, not to build on, until the licence is public.
Mistral Large 4 is worth a Canadian team’s attention, but only as a trial and not yet as a foundation. Mistral’s changelog shows the model in public preview from 6 October as an open-weight multimodal mixture-of-experts model with a 1 million token window, with weights promised soon. MarkTechPost reports 1.05 trillion total parameters and 49 billion active per token. That is the claim. What a business can rely on today is smaller.
TL;DR
- Attributed claim: API list prices of $1.36 per million input tokens and $4.18 per million output tokens, per MarkTechPost.
- Verified fact: Mistral’s changelog separately says launch pricing is 50% off for two weeks, so check which figure applies to you.
- Opinion: run a two-week test on your own documents, and do not commit to the open-weight route until the licence appears.
What do we know and not know?
We know the model exists, is available through Mistral’s API in preview and is described as open weight. We do not yet have the weights, the licence terms, the expert count or the architecture details. MarkTechPost says weights are promised by the end of October. Open weight describes access to the trained parameters, and licences vary on commercial use, so the word alone settles nothing for a company.
Who measured the benchmarks?
Mostly Mistral. The reported cybersecurity scores of 93% on Cybench and 82% on CyberGym-E2E, and the coding scores near 62% on DeepSWE, come from the vendor. The one outside number is a Surge AI human evaluation that placed the model second of five, at 3.74 out of 5, behind Claude Opus 5 at 4.22. That is a useful, modest result. Vendor benchmarks on tasks the vendor chose are a starting point and not a purchase case.
The Limitation
The limit is the gap between a preview and a product. A 1 million token window does not mean the model uses all of it well, and long-context accuracy usually drops with length, so test with your own longest files. A model trained in European data centres may suit firms wanting non-US hosting, but where your data is processed under the API is a contract question, not a spec sheet one. Our sovereign AI piece covers how to ask it.
What should you do in the next two weeks?
- Take ten real documents your staff handle, including your longest, and run them through the preview.
- Score answers against a checklist written before you start, not after.
- Ask Mistral in writing where the API processes and stores data, and for how long.
- Wait for the licence before any plan to self-host.
What does the sceptic say?
The sceptic says a frontier open-weight model from outside the US is exactly what Canadian firms want, so waiting is a mistake. Testing now costs little, and I agree with that. Building now does not. What would prove me wrong is a permissive commercial licence and independent evaluations that match the vendor numbers.
What to watch
Watch for the licence, the end of the 50% launch pricing, independent long-context tests and whether weights actually ship in October.
Frequently asked questions
Is Mistral Large 4 open source?
Mistral describes it as open weight, with weights promised soon. The licence has not been announced in the sources checked, so commercial terms are unknown.
How much does Mistral Large 4 cost?
MarkTechPost reports $1.36 per million input tokens and $4.18 per million output tokens. Mistral’s changelog says launch pricing is 50% off for two weeks, so confirm on its price page.
Can Canadian businesses use Mistral Large 4?
Yes through the API in preview. Ask Mistral where data is processed and stored before sending anything sensitive.
The decision in one line
Test Mistral Large 4 on your own documents for two weeks, and do not build on it until the licence is public.
Written by Priya Chen, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 7 October 2026, written and fact-checked on 7 October 2026. Sources are linked on the claims they support.
