Google ships EmbeddingGemma 2 under Apache 2.0

EmbeddingGemma 2: Google DeepMind released EmbeddingGemma 2 on Tuesday, a 740-million-parameter multimodal embedding model under Apache 2.0. It maps text, code, images, video, and audio into one 768-dimension space, Google said. Text-only workloads can run at as little as 270 million parameters; optional vision and audio encoders add about 170 million and about 300 million parameters for the full multimodal model, Google said. On a Pixel 11 Pro, quantized, Google says active RAM is about 191 MB for text-only weights and about 567 MB for the full model. MTEB Code rose to 78.68 from 68.76 versus the prior EmbeddingGemma, Google said. Context is 8,000 tokens. Weights are on Hugging Face and Kaggle.

OpenAI math: OpenAI on Tuesday published a GitHub set of mathematical results from an internal frontier model, with Lean formalizations and 10 reasoning summaries. The average result used about three hours of ChatGPT Pro-equivalent compute, OpenAI said. The company says it followed advice from the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. The model itself is not released.

Korea banks: Police probe data leaks at seven financial firms, the Korea Herald reported.