Google Unveils EmbeddingGemma 2, Its Most Capable On-Device AI Model Yet, as Reflection Launches 500B-Parameter Open-Source Beam for Cheaper Frontier Reasoning

Google has unveiled EmbeddingGemma 2, an open-source model it bills as its most capable yet for on-device AI, packing 740 million parameters (plus a 270-million-parameter text-only variant) and handling images, video, audio and code while running semantic search and retrieval locally. In a parallel move, New York startup Reflection launched Beam, an open-source model with over 500 billion parameters for coding, reasoning and agentic workloads that promises frontier reasoning up to four times cheaper than larger open-weight models like GLM-5.2, approaching Qwen 3.8-Max on coding and agentic tasks.

• Google unveils EmbeddingGemma 2, an open-source model it calls its most capable yet for on-device work, at 740 million parameters with a 270-million-parameter text-only variant, handling code, images, video and audio beyond text.
• The model performs on-device semantic search, routing and retrieval—finding specific voice memo clips and searching hours of audio from a text request—and is best-in-class for embedded models under 1 billion parameters.
• New York-based startup Reflection unveiled Beam on Monday, an open-source model with more than 500 billion parameters for coding, reasoning and agentic workloads, offering frontier reasoning up to four times cheaper than larger open-weight models like GLM-5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.