NVIDIA releases Nemotron-3-Embed-8B-BF16, a multilingual embedding model topping its RTEB benchmark

NVIDIA published Nemotron-3-Embed-8B-BF16 on Hugging Face on July 14, 2026, an open-weights text embedding model built for retrieval-augmented generation (RAG) and semantic search across 34 languages, released under NVIDIA's OpenMDW-1.1 open license.

What's new

Nemotron-3-Embed-8B-BF16 is, in NVIDIA's words, "a versatile text embedding model trained by NVIDIA and optimized for retrieval and semantic similarity tasks." The model is built on an 8-billion-parameter Ministral3 architecture, produces 4,096-dimensional embeddings, and supports sequences up to 32,768 tokens — long enough to embed full documents rather than short chunks.

On NVIDIA's own benchmark comparisons, the 8B model scores:

• 78.46 on RTEB-16 (retrieval)
• 60.60 on ViDoRE-V3 text
• 75.45 on MMTEB retrieval

That's a consistent step up from NVIDIA's smaller sibling release, Nemotron-3-Embed-1B-BF16 (72.38 / 57.74 / 71.04 on the same three benchmarks), with the largest gap — over 6 points — on RTEB-16. NVIDIA positions the model as suited to tasks like building multilingual question-answering systems over large text corpora using dense retrieval. Weights, config, and a sentence-transformers-compatible setup are published directly on the model's Hugging Face repository under the nvidia organization account.

Context

The release extends NVIDIA's Nemotron embedding lineup, which has increasingly focused on multilingual and long-context retrieval as RAG pipelines move from English-only prototypes into globally deployed products. NVIDIA has published several embedding variants in this family this year at different parameter counts and precisions, giving developers a size/accuracy tradeoff rather than a single fixed model, and has been actively contesting the top of public multilingual retrieval leaderboards with the broader Nemotron line.

Why it matters

Embedding models are unglamorous compared to flagship chat releases, but they're the retrieval backbone underneath most production RAG systems, and benchmark leadership here has direct commercial weight for NVIDIA's enterprise AI stack — every retrieval-heavy deployment running on NVIDIA infrastructure benefits from a stronger, openly licensed embedding option that doesn't require a third-party API call. Shipping it open-weight, rather than API-gated, also keeps NVIDIA competitive against open embedding releases from labs like Cohere and BAAI, where developers increasingly expect to self-host rather than pay per embedding call.