Mistral Just Previewed a 1.05 Trillion Parameter AI Model It Nicknamed 'Le Chonk'

Mistral AI put a 1.05 trillion parameter model into public preview on October 6, three weeks before it plans to hand over the full weights, and gave it a nickname that undersells what it built: Le Chonk.

A model so large its own makers gave it a joke name

The model, officially called Mistral Large 4, is a mixture of experts system, which means it does not switch on all 1.05 trillion parameters for every answer. Only around 49 billion of them activate for a given request, with the rest sitting in reserve as specialized experts the model calls on only when a question needs them. That design lets Mistral claim reasoning closer to a far larger dense model while keeping the computing cost closer to a 50 billion parameter one.

The French company trained it over roughly two months on 4,000 Nvidia Grace Blackwell chips running in European data centers. Mistral says the result is fluent in more than 160 languages and handles text, images and other input types in a single system, rather than bolting vision on as an afterthought.

Mistral is positioning Large 4 against the open weight models coming out of China, DeepSeek and Alibaba’s Qwen line chief among them, rather than against the closed systems from OpenAI, Anthropic or Google. The company says it leads other open weight systems on cybersecurity, finance and manufacturing benchmarks, though those figures come from Mistral’s own testing and have not been checked independently. Access during the preview runs through Mistral’s API at $1.36 per million input tokens and $4.18 per million output tokens, a premium price for a model being sold partly on the promise of openness. Developers, security teams and government bodies get roughly three weeks to put it through its paces before the full release on October 27.

Teaching more defenders to use the AI that finds decades old bugs

Anthropic folded its Project Glasswing cybersecurity initiative into a wider program on October 6, opening access to an AI system built to hunt software vulnerabilities to a much longer list of applicants.

Glasswing launched earlier this year with twelve founding partners, among them AWS, Apple, Cisco, Microsoft and Nvidia, giving a restricted set of critical infrastructure operators access to Claude Mythos, Anthropic’s most capable model for autonomous coding work. In that early testing the system found more than 10,000 high or critical severity vulnerabilities across partner codebases, including a 27 year old flaw in the OpenBSD operating system and a 16 year old bug in the widely used FFmpeg video software, both defects that had sat undetected through years of ordinary human review.

The expanded effort, now called the Cyber Verification Program, splits access into three tiers. Defense Access covers incident response, malware reverse engineering and vulnerability validation, and Anthropic says it aims to answer those applications within a few days. Red Team Access goes to organizations doing authorized penetration testing, and Specialized Access is reserved for safety critical systems such as flight control software and power grid management. Each tier includes Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1. Where the original program limited itself to a short list of critical infrastructure operators, the expanded one takes applications from security firms, open source maintainers and qualifying independent researchers as well, through a public portal on Anthropic’s site.

Google pays billions to keep old reactors running longer

Alphabet and Constellation Energy signed a pair of long term power agreements on October 6 that will send Google as much as 3.59 gigawatts of electricity, most of it from nuclear plants Constellation already owns.

The larger piece of the deal is a 20 year agreement covering 890 megawatts of new nuclear capacity, generated by upgrading equipment at 11 Constellation owned reactors spread across Illinois, Pennsylvania and New Jersey rather than by building new plants from scratch. Constellation says it will spend more than $4.3 billion on the work, with the first uprated unit expected online by 2028. A second, 15 year agreement adds another 2,700 megawatts of conventional power from Constellation’s wider fleet in PJM, the regional grid serving much of the mid Atlantic and parts of the Midwest.

The deal puts Google in the same position as Microsoft and Meta, both of which have already signed their own multibillion dollar nuclear agreements this year to keep aging reactors running longer rather than wait years for new ones to be built. Nuclear plants produce power around the clock regardless of weather, which has made them attractive to companies trying to lock in electricity for data centers that never stop drawing current. Google says the power from this deal will go toward expanding its AI data center footprint in the PJM region, where demand from new computing facilities has already strained the grid enough to draw scrutiny from state regulators.