NVIDIA open-sources Nemotron 3 Embed models — the 8B version tops a new benchmark

The models that power how AI systems search, retrieve, and cross-reference information don’t get much attention, but they’re quietly becoming one of the most competitive layers in the AI stack. Embedding models — the ones that turn text into vectors for similarity search — are the engine behind RAG pipelines, AI agent memory, and enterprise search. On Thursday, NVIDIA released an open-weight family called Nemotron 3 Embed that targets exactly this layer, and the largest variant is already topping a new benchmark from Hugging Face.

All three models come with a 32,000-token context window and are available under a commercial-friendly open-weight license. NVIDIA is hosting them on Hugging Face and through its NIM inference microservice, free to use. The company says the architecture is designed to reduce how often AI agents need to make redundant retrieval calls — cutting down on unnecessary inference triggers and lowering token consumption in the process.

The family splits into three variants for different deployment profiles:

Nemotron-3-Embed-8B-BF16 is the flagship, with 8 billion parameters. NVIDIA positions this one for enterprise applications where precision matters and the cost of a wrong retrieval is high. It’s built on top of Ministral-3-8B-Instruct-2512.

Nemotron-3-Embed-1B-BF16 is the efficiency play, with 1.14 billion parameters. This is distilled from a larger parent model based on Ministral-3-3B-Instruct-2512, and targets environments where latency and cost per query are the binding constraints.

Nemotron-3-Embed-1B-NVFP4 is the same 1B model, but optimized for NVIDIA’s Blackwell architecture using NVFP4 (NVIDIA’s 4-bit floating point format). It maintains the same retrieval accuracy as the 1B-BF16 version while cutting memory usage and boosting throughput by up to 2x.

On performance, the 8B model scored 78.5% on the RTEB benchmark — enough for first place — and 75.5% on the MMTEB retrieval task. RTEB (Retrieval Embedding Benchmark) is a relatively new evaluation suite from the Hugging Face team and the MTEB community, designed specifically to measure how well embedding models perform in real-world RAG and agent scenarios, rather than on synthetic or overly narrow tasks.

The open-weight strategy here is worth noting. Other embedding models from companies like OpenAI and Cohere remain proprietary and API-only. By releasing Nemotron 3 Embed under a permissive license with both a high-end and a lightweight variant, NVIDIA is giving developers a self-hosted option that can run on their own hardware — and making it easy to find the right cost-accuracy tradeoff without vendor lock-in.