EmbeddingGemma 2 dim tradeoff: 128d shrinks vectors 6x but MMEB drops to 45.65
tomaarsen · x · 2026-10-07
tomaarsen tested EmbeddingGemma 2's Matryoshka truncation: 128d vectors are 6x smaller, but MMEB v2 overall drops from 59.01 at 768d to 45.65. The model card positions 128d for text-only workloads; at 256d the quality impact is much smaller for mixed modalities — a third of the storage with MTEB multilingual v2 only dropping 61.36→60.41. In Sentence Transformers: set truncatedim=256, normalizeembeddings=True, and re-normalize after truncating.
More from Models
- llama.cpp ships Day-0 support for Google's EmbeddingGemma 2 — ggerganov · 2026-10-07
- Trying to Plug Open-Source Mistral Into an Agentic Coder Just Doesn't Work, Says Berman — MatthewBerman · 2026-10-07
- antirez: DeepSeek v4.1 outscores Mistral Large 4 on DeepSWE 1.1 and other benchmarks — antirez · 2026-10-07
- StartLux claims its 27B model beats Jev AI on 31 of 38 benchmarks, self-reported results — Dr_Singularity · 2026-10-07
- 24 models tested on 669 clinical decisions: Jev stays #1 as two free models close in — MaziyarPanahi · 2026-10-07
- Never run EmbeddingGemma 2 in FP16: use BF16 or FP32 to avoid NaNs — tomaarsen · 2026-10-07