← Back to live feed1 story
Google released EmbeddingGemma 2, a 740M parameter multimodal model that benchmarks comparably to OpenAI's embedding-3-large and Qwen 8B while running on consumer CPUs rather than GPUs. The model has become the 4th most liked in the company's history after surpassing EmbeddingGemma 1 in popularity. Developers have used the model to build DigUp, an open-source tool for Mac that indexes local files and podcasts, with Google employing Matryoshka embeddings to store only the first 256 of 768 values in a vector, reducing storage needs to 1/3 while maintaining about 95% of the retrieval quality for visuals and speech.
Sign in to suggest
Key sources
- SOURCE@dorkitude“comparable to open AI's embedding-3-large and Qwen 8B, which are two of the most common GPU models out there”x.com
- SOURCE@tomaarsen“surpasses EG1's likes to become the 4th most liked Google model ever, as well as the 3rd most liked Sentence Transformers model ever”x.com
- SUPPORT@learnopencv“a video and its reverse are 99%+ similar”x.com
- SUPPORT@tomaarsen“unlike LLMs, you can't just easily switch if the host deprecates the model, you have to reindex all docs”x.com
- SUPPORT@theturingpost“search videos with a spoken description, audio with a question, and images with a sketch - all on your device, with just 740M parameters”x.com
- SUPPORT@googlegemma“DigUp runs @GoogleDeepMind's new EmbeddingGemma 2 on-device: describe a video moment, a podcast minute, a PDF page, a photo or code”x.com
- SOURCEmarketbrief.now