Mistral Previews 1T-Parameter Le Chonk Model; Google, Meta, Anthropic Announce New AI Tools

Mistral Large 4 Enters Public Preview

Mistral launched a public preview of Mistral Large 4, a one trillion parameter open weight model with 49 billion active parameters that the company says is the strongest open weight model developed in the US or Europe on aggregated benchmarks. The model, nicknamed Le Chonk, is natively multimodal and was trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own data centers in Europe, including its cluster in Bruyères-le-Châtel.

Mistral says the model is state of the art on workloads including cyber defense, manufacturing and finance, and that it surpasses closed frontier models on visual grounding. On one security test that asks a model to reproduce a real vulnerability in open source software and then patch it, ML4 scored 82%, the highest of any model. The model also scored 62% on DeepSWE v1.1. Open weights are scheduled for release on October 27.

Google DeepMind Releases EmbeddingGemma 2

Google DeepMind launched EmbeddingGemma 2, a 740 million parameter open model that maps text, code, images, video and audio into a single shared embedding space. Released under an Apache 2.0 license, the model is designed for on-device use and combines a 270 million parameter text model with vision and audio encoders. Google says it uses about 191MB of RAM for text only and about 567MB for full multimodal use, measured on a Pixel 11 Pro.

Industry Groups Back Personal Agent Protocol

Meta, Sierra, Walmart, Instinct, Shopify, Stripe, Genesys and RocketOTD published the Personal Agent Protocol, an open standard intended to define how personal AI agents interact with businesses. The announcement follows the launch of Meta's personal agent Muse. Amazon was notably absent from the group. Separately, Stripe said agentic commerce hit an inflection point in recent weeks with the rise of personal agents.

Anthropic Expands Cyber Verification Program

Anthropic integrated its Cyber Verification Program and Project Glasswing into an expanded version of CVP with three access tiers. The company said it found more than 5,500 verified vulnerabilities between April and October, while Glasswing partners found more than 129,000 between April and July, including over 33,000 rated critical or high severity.

Sources: Techmeme