Model Release

NVIDIA Nemotron 3 Embed Takes the Top RTEB Retrieval Slot

By Kaleido Field Staff · July 17, 2026

What happened and why it matters

The release matters to answer systems because retrieval quality determines which passages an agent sees before it writes, but leaderboard gains do not settle latency, cost or domain fit.

Primary source

Primary reference: NVIDIA on Hugging Face: Nemotron 3 Embed Wins RTEB. Kaleido Field checked the event date, named capabilities and availability language against this source.

What was released

The collection contains an 8B BF16 model, a 1B BF16 model and a 1B NVFP4 variant. NVIDIA provides open weights, a 32,000-token context window, multilingual and code retrieval support, plus fine-tuning and distillation recipes.

Deployment options include Hugging Face, NVIDIA NIM and several inference partners.

The benchmark evidence

NVIDIA reports 78.5 on RTEB and 75.5 on MMTEB Retrieval for Nemotron-3-Embed-8B-BF16. The linked RTEB leaderboard showed the model in first place in NVIDIA's July 15 snapshot.

The 1B BF16 model is reported at 72.4 on RTEB and 71.0 on MMTEB Retrieval. These values should stay attached to the named model, metric and source date.

What still needs testing

A public retrieval score does not establish the best model for every corpus. Production teams still need to measure indexing cost, query latency, memory use, language coverage and retrieval quality on their own documents.

Partner testimonials and fine-tuning gains in the release are vendor-selected evidence. They are useful leads for testing, not independent proof of universal superiority.

Evidence boundary

This page reports a dated event from a named primary source. Company specifications and adoption statements remain attributed claims unless independent evidence is cited above.