Two very different visions of how powerful artificial intelligence should be shared collided this week. On Thursday, Paris-based Mistral AI released the full weights of Mistral Large 4, a model with about one trillion parameters that the company has nicknamed “Le Chonk”. A day earlier, Google unveiled Gemini 4 Carbon, its most capable model yet, but said only a select group of cyber defenders can use it for now.
What Mistral has released
Releasing the weights means anyone can download the model itself, rather than only renting access to it over the internet. Businesses and governments can run it on their own servers, fine-tune it and build products on it. TechCrunch reported that the weights are on Hugging Face as well as through Mistral’s own service, and Mistral says it is also available on platforms including OpenRouter, NVIDIA’s services and Microsoft’s Azure Foundry.
The model, first shown as a public preview earlier in October, is a so-called mixture-of-experts design. According to Mistral, it handles instructions, step-by-step reasoning and agent-style tasks in one system, accepts images as well as text, and works in more than 160 languages. Mistral says it was trained from scratch on European infrastructure.
It is released under what Mistral calls a modified MIT licence. That allows free use, modification and commercial use, with one big exception: companies that want to sell it as a hosted AI service need a separate commercial licence from Mistral. For those paying Mistral directly, prices start at $0.70 per million input tokens and $2.10 per million output tokens.
The cybersecurity pitch
Mistral is selling the model hard as a defensive cyber tool. TechCrunch reported that Pierre Stock, Mistral’s vice-president of science, argued that defenders need the same capabilities as attackers quickly, and that an open-weight model cannot have its access cut off. Chief executive Arthur Mensch said AI should “be a tool for empowerment, not dominance”.
Mistral’s own figures claim the model leads all open models on an independent coding-agent ranking from Artificial Analysis, matches the Chinese open model GLM-5.3 while using 47% fewer tokens per task, and refuses malicious cyber requests 96.6% of the time on average. Those benchmark numbers come from Mistral and have not been independently verified by us. Mistral also places the model behind Anthropic’s Claude Opus 5.5 in one blind human-preference coding test.
Google takes the opposite route
Google’s Gemini 4 Carbon, announced on Wednesday alongside Alphabet’s quarterly results, tops the general Artificial Analysis Intelligence Index with a score of 59, just ahead of Claude Opus 5.5 on 58 and OpenAI’s GPT-6 Astra on 57, according to The Next Web.
But Google is not opening the doors. Carbon is going first to a subset of cyber defenders, with more trusted partners to be added through its Fairwind Program, and only later to developers, businesses and consumers, with no date given. Google says Carbon is very strong at finding, exploiting and fixing software vulnerabilities, and it is taking part in the US government’s voluntary pre-release testing process. The Next Web also noted VentureBeat’s report that Google has not yet published a full technical report, and that Carbon’s hallucination rate on one knowledge test rose compared with its predecessor.
The contrast is stark. Google is treating cyber capability as something to ration carefully. Mistral is betting that giving it away arms defenders faster than it helps attackers. Critics of open release have long warned that once weights are public, nobody can recall them or stop misuse.
What happens next
For Europe, Mistral’s release is a big moment in the push for so-called sovereign AI: models that EU governments and companies can control without depending on US cloud giants. Expect security teams and public bodies to start testing it in the coming weeks, and expect independent researchers to put Mistral’s safety and refusal claims under the microscope.
Google, meanwhile, says wider access to Carbon will come “as soon as possible”. Its launch price of $2 per million input tokens and $10 per million output is set to rise later to $4 and $20, according to The Next Web, though Google has not said when.
Sources
- TechCrunch: Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity (29 October 2026)
- Mistral AI: Mistral Large 4 (Le Chonk) is now open weight (October 2026)
- The Next Web: Google’s Gemini 4 Carbon tops AI rankings, but only cyber defenders get it (29 October 2026)
Featured image: photo by Igor Omilaev on Unsplash.
Leave a Reply