If you are budgeting for RAG or agentic retrieval, two numbers decide what it really costs you — and neither appears on the leaderboard you are being shown.
The licence problem
Retrieval research has a de facto anchor model: NV-Embed-v2. It is the point of comparison that leading multi-hop systems are measured against. It is also licensed cc-by-nc-4.0 — non-commercial.
Three of the four leading MuSiQue systems depend on it, and none of them disclose that. If you read those results as a buying signal and build on the same foundation, you inherit a licence that does not permit commercial deployment. That is a procurement problem discovered late, after the architecture is set.
Until mid-2026, commercially-licensed models genuinely did lag — by 2.31 Recall@5 points. That gap has closed. NVIDIA's Nemotron-3-Embed-8B matches the anchor — it does not beat it — and is the only entrant that is commercially licensed, free to self-host, and statistically indistinguishable from the model the field benchmarks against.
Put plainly: the quality penalty for choosing something you are allowed to deploy, on your own hardware, inside your own perimeter, is now effectively zero. That was not true a year ago. Most procurement assumptions still price in a penalty that no longer exists.
The cost that is never quoted
Benchmarks report accuracy. They do not report what it costs to build the index in the first place — or to rebuild it, which you do every time your corpus changes.
An eleven-fold spread on the same task is not a rounding difference — it is the difference between a line item and a budget problem. And because it recurs on every re-index, it compounds with how often your data moves, which for most real deployments is continuously.
Why this matters if you are the one signing
The question that gets asked is “which model is smartest.” On a public benchmark the largest frontier model usually wins — partly because the benchmark is built from material it has already absorbed. That is an in-sample score.
On your documents — the ones it has never seen, the ones that cannot leave your building — that advantage narrows sharply. What remains is the ability to read the evidence actually in front of it, and a self-hosted model you run on your own hardware does that job.
So the two questions worth asking a vendor are not about leaderboards:
1. Am I licensed to deploy this where my data lives?
2. What does it cost me every time my data changes?
The paper is the evidence for both: thirteen embedders from eight makers, evaluated on MuSiQue through one identical harness, with bootstrap confidence intervals rather than single-point scores.
The follow-up takes this from benchmark to deployment — what these findings cost, and what they change, for organisations that cannot send their data anywhere. If you want it when it publishes, ask and you will get it.
Email me when Part 2 publishesGoes to the author directly. No list, no tracking pixel in the mail, no forwarding — and nothing else is ever sent.
Read the full paper: arxiv.org/abs/2608.16096