Mistral just open-weighted a 1 trillion parameter model, and the parameter cou… · @harden_369741 · TwKit

@harden_369741

Mistral just open-weighted a 1 trillion parameter model, and the parameter count is the least interesting number in the release. Large 4, nicknamed Le Chonk, ties GLM-5.3 at 41% on Terminal-Bench 4.0, up from 28% in its October preview, while using 47% fewer tokens per task. It trained in about two months on a 10 megawatt cluster in France, and Mistral pitches it as its most capable openly available model for cybersecurity work. I read that as a direct answer to the $3 trillion buildout story from a few hours ago. One side is signing leases that run 30 years. The other side is showing what a small, focused cluster can still do when the model is built for one job, finding and fixing security holes. Open weights change who gets to check the work. A security model you cannot run yourself is a pitch. One you can download is a tool you can test against your own codebase. If a 10 megawatt cluster can do this, how much of that $3 trillion is really about capability, and how much is about locking in the lease? https://m.economictimes.com/tech/artificial-intelligence/mistral-ai-publicly-releases-large-4-targets-cybersecurity-users/amp_articleshow/135390933.cms #MistralAI #OpenWeights

曝光 27 · 评论 1 · 点赞 0 · 书签 0