Hugging Face engineer Erik Kaum introduced Lattice, a static embedding model built on tokenization and a lookup table with mean pooling and L2 normalization rather than transformer attention layers. The model was trained in two stages on 660 million query-document pairs (contrastive pre-training followed by hard-negative fine-tuning), reaching 0.4749 NDCG@10 on decontaminated BEIR benchmarks. Using int4-row quantization at 512 dimensions, its weights compress to just 7.94 MB, and a pure-Rust runtime processes 9.52 million tokens per second on consumer hardware. The team used it to embed all 6.4 million English Wikipedia articles in about seven minutes and 26 seconds.
