Lattice embeds Wikipedia in seven minutes

Lattice, a static embedding model trained on 660 million query-document pairs, embedded all 6.4 million English Wikipedia articles in 7 minutes 26 seconds on an eight-core M2 MacBook Air. Its creator’s Hugging Face technical report says the int4-row, 512-dimension artifact occupies 7.94 MB and scores 0.4697 NDCG@10, effectively matching the same-width full-precision model despite static embeddings’ known limits with word order and polysemy. For teams evaluating retrieval systems after Onton’s benchmark earned a pilot rather than a switch, the result makes a tiny CPU retriever worth filing away as a first-stage candidate generator—not a replacement for a reranker.