Nomic Embed: Training a Reproducible Long Context Text Embedder

Open embedding models could not be reproduced because training data and code were withheld, and long-context embedders were weak.

Matryoshka-trained long-context embedder (learns to prefix-truncate to any shorter context length) with LoRA adapters applied to a causal LLM; the full curated training data and code are released so the model can be replicated end to end. Long inputs are truncated from the middle, not the end.

Claims to outperform OpenAI ada-002 and text-embedding-3-small on both the short-context MTEB benchmark and the long-context LoCo benchmark, at 8192 context length. Vendor self-reported comparison against closed APIs with no cost or latency disclosure - this is a quality claim only.

Venue TMLR (Semantic Scholar records 'Trans. Mach. Learn. Res.'; the arXiv comment links the OpenReview forum IPmzyQSiQE). The schema has no TMLR entry so 'other' is used. THE MECHANISM WORTH STEALING IS MATRYOSHKA + MIDDLE-TRUNCATION: a single checkpoint serves every context length by truncation, which means one set of weights serves short-query and long-document embeddings instead of two models. For serving that halves resident weight memory per embedding index. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as TMLR (same). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. VENUE RECLASSIFIED 2026-10-10 (crawler/w11-venue-enum): venue 'other' -> 'tmlr', venue_kind 'journal'. Confirmed by record notes and agreeing publisher/index deposit in data/sources/.