Better Evidence, Better Agents: NVIDIA Nemotron Retriever on FriendliAI, Available Day-0

Starting today, developers can run NVIDIA Nemotron 3.5 Embed 1B, Nemotron 3.5 Embed 8B, and Nemotron 3.5 Rerank 8B on Friendli Dedicated Endpoints. Part of the NVIDIA Nemotron Retriever collection, these open models give agents a multimodal retrieval layer: embeddings that understand both text and visuals, and a reranker that hands the reasoning model only the strongest evidence.

Enterprise agents are only as good as the information they can retrieve

As companies move agents from demos to real work, retrieval becomes the step everything else depends on. Enterprise corpora span unstructured documents, images, tables, and complex multimodal and multilingual content, and agents can incur significant inference costs to parse, retrieve, and reason through all of it.

When an agent retrieves the wrong information or misses relevant context, it produces unreliable answers, responds slowly, or fails to complete the workflow. Many teams still rely on legacy search, limited retrieval models, and brittle pipelines that don't scale across real enterprise data. The result: agents that spend too many tokens while coming to incorrect conclusions.

Agents need a reliable, cost-efficient way to find the right information across every source, including answers that live in a chart or a scanned page rather than in plain text.

What is NVIDIA Nemotron Retriever?

NVIDIA Nemotron Retriever is a collection of open, leaderboard-winning models for building world-class retrieval systems for agents. Teams can use the models out of the box, post-train them for their domain, and integrate them into their own agent solutions to improve retrieval quality and deliver more reliable agent outcomes.

FriendliAI supports the two models that sit at the center of the retrieval loop: Embed, which makes enterprise content searchable, and Rerank, which moves the most relevant evidence to the top.

What you can build

• Agentic RAG over visual documents — retrieve answers from tables, charts, diagrams, and full document pages, such as financial reports, technical manuals, and slide decks, without losing what lives outside the plain text.
• Multilingual enterprise search — ground copilots and assistants in knowledge bases that span languages and formats, with one embedding space for text and visual content.
• Iterative, multi-source agent retrieval — support agents that search repeatedly, compare evidence across sources, and use retrieval as memory, with Rerank deciding what is worth passing to the reasoning model.
• Domain-specific retrieval — post-train Embed and Rerank on your own terminology and relevance criteria for legal, financial, healthcare, or internal enterprise data.

The common thread: high-frequency retrieval on the critical path of every agent turn, where accuracy decides the answer and efficiency decides the bill.

Running Nemotron Retriever on FriendliAI

Retrieval is where efficiency compounds. A single agent task can fan out into dozens of embedding and reranking calls: every query rewrite, memory lookup, and evidence check is another inference request. FriendliAI's inference stack is built to make exactly this kind of high-frequency, latency-sensitive workload fast and economical in production.

What FriendliAI brings to Nemotron Retriever:

High-throughput retrieval inference. FriendliAI's inference engine maximizes throughput per GPU through continuous batching and kernel-level optimizations, well suited to the bursty, high-volume query patterns of agentic retrieval.

Low latency on the critical path. Retrieval sits in front of every reasoning step, so its latency adds up across the agent loop. Fast embedding and reranking keep each turn moving.

Cost-efficient retrieval at scale. Retrieval only becomes a default agent capability when it's inexpensive enough to call constantly. FriendliAI delivers higher tokens-per-dollar than alternative OSS serving stacks, making it practical to embed and rerank as often as agents need.

Production-grade reliability. Dedicated endpoints, auto-scaling, operational monitoring, and a 99.99% uptime SLA, because an agent workflow fails when any step fails.

Model and data control. Open weights plus dedicated GPU capacity give teams control over the models, data handling, and deployment, including customized versions post-trained for their own domain.

Get started

Nemotron 3.5 Embed 1B, Embed 8B, and Rerank 8B are available on Friendli Dedicated Endpoints starting today: reserved GPU capacity with predictable latency for production retrieval workloads.

1️⃣ Navigate to the dedicated endpoint creation page.

2️⃣ Choose your model: nvidia/Nemotron-3.5-Embed-1B-BF16, nvidia/Nemotron-3.5-Embed-8B-BF16, or nvidia/Nemotron-3.5-Rerank-8B-BF16

3️⃣ Click "Create"

Prerequisites

• A FriendliAI account
• A Friendli API Key from Friendli Suite settings

Install the package

pip install friendli

Environment Setup

export FRIENDLI_API_KEY="your-api-key-here"

Example Code (Embed)

import os
from friendli import SyncFriendli

with SyncFriendli(
    token=os.environ["FRIENDLI_API_KEY"],
) as friendli:
    res = friendli.dedicated.embeddings.embeddings(
        model="your-dedicated-endpoint-id",
        input_="Hello, Nemotron 3.5 Embed!",
        encoding_format="float",
    )

    # Handle response
    print(res.data[0].embedding)

For teams with specialized terminology or relevance criteria, the open weights and post-training recipes provide a clear path to customize Embed and Rerank for your own data, and deploy the customized versions on FriendliAI.

Deploy NVIDIA Nemotron at Scale with FriendliAI

FriendliAI is proud to partner with NVIDIA to bring the Nemotron family of models to developers and businesses building production-ready, agentic AI systems.

NVIDIA Nemotron Retriever gives agents better evidence: multimodal embeddings that make every page searchable, and a reranker that hands the reasoning model only what matters. FriendliAI is your platform to take it to production.

👉 Launch your Nemotron Retriever Dedicated Endpoint today and give your agents a retrieval layer that keeps up with them.