DeepInfra Now Serves NVIDIA Nemotron 3 Embed: Frontier Retrieval for RAG and Agents

Agents live or die on retrieval. Miss the right passage, code block, or document, and the agent reasons from the wrong context — wasting tokens and reducing answer quality. And agents retrieve constantly: decomposing tasks, rewriting queries, searching memory, inspecting code. For retrieval to be the default for every agent decision, the embedding model has to be accurate, fast, and cost-effective, all at once.

That's exactly what NVIDIA Nemotron 3 Embed delivers — and it's now available on DeepInfra.

The retrieval tradeoff, solved

Enterprise retrieval usually forces tradeoffs: better accuracy means paying more per query, larger models are harder to serve, and fast answers can mean missing the right content. Nemotron 3 Embed comes in two sizes so you don't have to pick one corner of that triangle:

• Nemotron 3 Embed 8B: frontier retrieval accuracy, topping the RTEB leaderboard across open and closed embedding models. Built for applications where maximum retrieval quality matters most.
• Nemotron 3 Embed 1B: an efficient production model that retains more than 95% of the 8B model's retrieval accuracy while dramatically reducing serving cost. Available in both BF16 and NVFP4 — the NVFP4 variant is optimized for NVIDIA Blackwell GPUs, delivering up to 2x higher throughput while retaining 99% of BF16 retrieval accuracy.

Whether you're optimizing for maximum retrieval quality or production-scale throughput, NVIDIA Nemotron 3 Embed provides the right model.

Why it matters for agents

Multi-turn agents retrieve repeatedly — for planning, long-term memory, code understanding, tool use, and multi-step reasoning. Weak retrieval means more turns, more tokens, and more hallucination risk. Strong retrieval reduces irrelevant context and keeps agents grounded. Nemotron 3 Embed is a stronger retrieval layer for agentic retrieval, query decomposition, query rewriting, code retrieval, enterprise search, and RAG applications.

Open, and production-ready

Nemotron 3 Embed ships with open weights, datasets, and recipes, so you can inspect it, tune it, and fine-tune it for your domain. No black box, no lock-in. On DeepInfra, both sizes are available now through our standard OpenAI-compatible API, so you can move from experimentation to production without changing infrastructure.

Get started

Generating embeddings requires only a single OpenAI-compatible API call:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepinfra.com/v1/openai",
    api_key="$DEEPINFRA_TOKEN",
)

resp = client.embeddings.create(
    model="nvidia/Nemotron-3-Embed-8B",
    input=["How do I reset my password?"],
)

print(resp.data[0].embedding[:8])
copy

For high-throughput workloads, simply swap in nvidia/Nemotron-3-Embed-1B-BF16, or nvidia/Nemotron-3-Embed-1B-NVFP4 for maximum throughput on NVIDIA Blackwell GPUs. All three models are available on DeepInfra today.

• Explore the models: Nemotron 3 Embed 8B, 1B BF16, 1B NVFP4
• Embeddings API docs: docs.deepinfra.com/apis/embeddings

Have questions or need help? Reach out at feedback@deepinfra.com, join our Discord, or connect with us on X (@DeepInfra) — we're happy to help.