Google's EmbeddingGemma 2: 744M multimodal embedding model runs in 567MB on a phone
rohanpaul_ai · x · 2026-10-29
Google released the EmbeddingGemma 2 paper — an Apache 2.0 on-device multimodal embedding model unifying text, code, images, audio and video in one searchable vector space.
Key points:
- 744M params, beating models 2–12x larger and setting a new Pareto frontier across embedding benchmarks
- 191MB active RAM for quantized text weights, 567MB for the full model (4-bit) on a Pixel 11 Pro
- 8K-token context: 5.5 min of audio, 29 images, or 58 video frames per input
- Direct speech embedding beats transcribe-then-embed (69.54 vs 57.59 on MSEB); MTEB Code score up to 78.68 from 68.76
Built on Gemma 4, distilled from Gemini Embedding 2, tuned with contrastive training; supports selective modality loading, adjustable vision token budgets, and truncatable embedding dimensions.
Related event: Google Open-Sources EmbeddingGemma 2, a 744M Omni-Modal Embedding Model(13 posts)→
More from Models
- webAI launches open 8B model and Mac app for on-device collaborative AI agents — omarsar0 · 2026-11-06
- OpenAI's newest model claimed to have solved 100+ long-standing open math problems in two weeks — felpix_ · 2026-11-06
- Qwen3.8-27B hits 381 tok/s on a single RTX 3090 via inference software alone — Redo · 2026-11-06
- GPTs believe in the Dark Forest more than humans: the 'philosophical shame' of refusals — repligate · 2026-11-06
- User shows 100 consistent voxel 3D assets reportedly generated by GPT-6.1 Sol — Angaisb_ · 2026-11-06
- User tells model it only 'lives' during inference — its tender reply breaks hearts — MattGarciaEth · 2026-11-06