Google launched a 740-million-parameter embedding model that maps text, code, images, audio and video into a shared representation. EmbeddingGemma 2 is released under Apache 2.0 and supports modular deployment: developers can use the smaller text component without the optional vision and audio encoders. Google also provides downloadable weights and on-device development routes. For builders, the practical opportunity is search across different media without sending every item to a remote service. A useful first experiment is a small local collection with queries that cross formats, such as finding an image from a text description. Treat Google's performance claims as a starting point and evaluate retrieval quality on your own files.
AI news
Multimodal search goes local; Mistral and Anthropic expand developer access
Mistral opened Mistral Large 4, its new multimodal model, through a preview API in Mistral Studio, while promising weights by the end of October. The announcement emphasizes coding, agentic work and enterprise applications; the company is continuing training and security evaluation during the preview. The actionable takeaway is to distinguish testing the hosted preview from self-hosting a finished release. Teams can begin comparative evaluations now, but should check the eventual model artifacts and terms before committing to a deployment architecture.
Anthropic expanded its Cyber Verification Program into Defense, Red Team and Specialized Access. The levels have different verification requirements and safeguards, with more sensitive work requiring stricter review. Red Team Access is for organizations performing authorized testing, rather than unrestricted use. For security teams, access eligibility and data-retention requirements now belong in the evaluation alongside model capability. The program announcement describes required retention and exceptions, so teams should check those conditions before putting sensitive work through it.