Google's EmbeddingGemma 2 unifies text, image, video embeddings; 1-bit quantization keeps 95% of search quality
solyarisoftware · x · 2026-10-12
- EmbeddingGemma 2 launched: a 740M-parameter lightweight multimodal embedding model from Google that maps text, code, images, video, and audio into one unified embedding space. Apache 2.0 licensed, optimized for on-device use, with an 8K context window (4x the text-only version) and flexible dimensions from 768 down to 128 via Matryoshka Representation Learning.
- Pastel Labs' photo anchor trick: they published an averaged photo vector as a spatial anchor — subtract it before quantizing to 1 or 4 bits. On COCO at 768 dims, 4-bit kept 0.997 of full-precision search quality; 1-bit kept 0.954, or 0.998 after reranking the top 20.
- Storage: 96 bytes per photo at 1-bit vs 3,072 bytes for float32.
Related event: Google Open-Sources EmbeddingGemma 2, a Multimodal Embedding Model(2 posts)→
More from Models
- Claude reportedly planned and coded a full video production in 2.5 hours — ZeroStateReflex · 2026-10-12
- KV caching explained: why it shapes GPU memory and drove the shift to GQA and MLA — blaizedsouza · 2026-10-12
- Claude Opus 5.5 vs GPT-6.1 Sol hands-on: this time GPT wins the vibe-check — CodeByPoonam · 2026-10-12
- DeepMind's AlphaProof Nexus reportedly cracks 9 Erdős problems, 44 OEIS conjectures — kimmonismus · 2026-10-12
- Codex keeps stalling on legit debugging over cybersecurity guardrails, user complains — Al_Grigor · 2026-10-12
- Third-party provider adds vision to Kimi K3, enables DeepSeek max thinking mode — sarahdrinkwater · 2026-10-12