Matryoshka training lets embeddings truncate to 256d with a third of the storage
tomaarsen · x · 2026-10-07
Google's new multimodal embedding model is Matryoshka-trained, supporting 768/512/256/128 dimensions. At 256d, vectors take a third of the storage while MTEB multilingual v2 only drops from 61.36 to 60.41. Usage in Sentence-Transformers: model.encode(..., truncatedim=256, normalizeembeddings=True), re-normalizing after truncation. Official 768d full-precision results: 67.84 NDCG@5 on MMEB v2 visual documents, 50.67 Hit@1 on video, and 69.54 MRR@10 on MSEB audio retrieval.
More from Multimodal
- Marc Andreessen boosts AI film contest SLOPTOBERFEST grand prize to $25,000 — zealcaiden · 2026-10-07
- Image generation pricing leak: $0.05 per 2K image, $0.076 per 4K — op7418 · 2026-10-07
- Live human votes plugged into Flow-GRPO to stop image models gaming reward models — lmoroney · 2026-10-07
- Open-source Tesseract converters let Claude natively edit Premiere and After Effects projects — LinusEkenstam · 2026-10-07
- DJ app's MCP server lets Claude drive real synths and drums instead of generating audio — tech_sand · 2026-10-07
- Spira launches Maxima 1.0, a text-to-social-video model post-trained on trending data — JaynitMakwana · 2026-10-07