A new audit of multi-hop retrieval benchmarks reveals that NVIDIA's Nemotron-3-Embed-8B has closed the performance gap between commercially licensed embedders and the non-commercial NV-Embed-v2 anchor on the MuSiQue benchmark. Released on July 16, 2026, this model is the only commercially licensed, free-to-self-host option that performs indistinguishably from the academic standard.

  • The study audits four leading MuSiQue systems (HippoRAG-2, PropRAG, SAG, KET-RAG), noting that three depend on the non-commercial NV-Embed-v2 without disclosure.
  • Until mid-2026, the best commercially licensed embedder trailed the anchor by 2.31 Recall@5 points; Nemotron-3-Embed now matches the anchor with a statistically insignificant difference of +0.24 at Recall@5.
  • Every other commercially licensed and free-to-self-host entrant sits 5.2 to 14.6 points below the anchor.
  • The audit highlights a significant cost disparity: three of five audited systems do not disclose indexing costs, with published GraphRAG figures spanning an 11x range (USD 2.30 vs USD 24.94 for a 5.64 MB corpus).
  • Extrapolated to 1 TB, this undisclosed choice separates costs from roughly USD 428K to $4.6M, with graph construction costs dwarfing embedding and answering expenses.

The findings emphasize that buyers must verify both licensing terms and indexing costs before relying on published benchmark numbers, as the paid-versus-free divide significantly impacts deployment viability.