The "Commercial Tax" of Multi-Hop Retrieval Benchmarks: Licensing and Cost Blind Spots

_reachsumit · x · 2026-08-18

The paper audits top multi-hop retrieval systems and finds their best results rely on embedding models with non-commercial licenses (e.g., NV-Embed-v2), while index construction costs are usually undisclosed. Comparing 13 embedding models, the study notes that as of mid-2026 the best commercially licensed model still trails the non-commercial anchor model by about 2.31 Recall@5 points. NVIDIA's release of Nemotron-3-Embed-8B narrowed this gap, making it the only self-hostable model with comparable performance. Moreover, API-based embedding models incur costs on every re-indexing, whereas self-hosted models carry no such overhead.

Original post →

More from Research

Research channel →