Deploy Multimodal AI Agents with NVIDIA Nemotron 3.5 Super VL on FriendliAI

Starting today, developers can run NVIDIA Nemotron 3.5 Super VL on FriendliAI Dedicated Endpoints. It brings visual understanding to agentic workflows and works alongside Nemotron 3.5 Lightning for high-volume text tasks and Nemotron 3 Ultra for complex reasoning and planning. Developers can choose the model that best fits each workload, rather than using the same model for every step.

Agents need to see, not just read

Always-on agents increasingly operate on visual information. A customer uploads a photo of a damaged product. A compliance workflow starts from a PDF full of tables and charts. A browsing agent lands on a page it has to interpret before it can click. In each case the model has to combine multimodal understanding with high-accuracy execution: reasoning across what it sees, calling the right tools, and preserving task-relevant context across many turns.

Stitching that together from separate vision, OCR, and text models adds latency and drops context at every seam. Nemotron 3.5 Super VL is built to remove those seams.

What is NVIDIA Nemotron 3.5 Super VL?

NVIDIA Nemotron 3.5 Super VL is an open multimodal model for long-running agents that need to see, reason, and act. It combines text, image, document, and video understanding with agentic reasoning and tool use in one model, so an agent can interpret visual input, reason across it, and call tools without handing off to a separate perception model.

Built on a 120B-parameter mixture-of-experts architecture with 12B active parameters, Super VL balances accuracy and efficiency for specialized, always-on agent tasks. Organizations can post-train it with open libraries using their own data, tools, policies, and workflows.

Within the Nemotron model family, Nemotron 3.5 Super VL is designed for complex enterprise tasks that require reasoning, tool use, and text and visual understanding. It sits alongside Nemotron 3.5 Lightning for high-volume text execution and Nemotron 3 Ultra for the hardest text-only reasoning. Route visual and document steps to Super VL; keep the rest on the model best suited to the job.

Why Nemotron 3.5 Super VL stands out

• Agentic execution with native visual understanding — High-accuracy reasoning and tool use combined with native document, image, and video understanding in one model, preserving task-relevant context throughout specialized workflows
• Highest accuracy for specialized domains — Frontier accuracy for agentic tasks in its class out of the box, and a clear path to post-train with NVIDIA NeMo on your own data, tools, and policies for the highest accuracy on specialized multimodal agent tasks
• Highest efficiency in its class — A hybrid mixture-of-experts architecture and multi-token prediction generate output faster. NVFP4 KV cache quantization shrinks the KV cache HBM footprint, creating headroom for more concurrent long-context sessions. In multi-turn agent workflows, retaining more cached prefixes means less KV cache eviction and recomputation, which supports higher throughput
• Model control — An open model you can customize and deploy across local, data center, and cloud environments while keeping control of model behavior, data handling, and agent workflows

Under the hood

Where Nemotron 3.5 Super VL fits/ Use Cases

Nemotron 3.5 Super VL earns its place wherever an agent's next step depends on something visual:

• Customer service — Resolve multi-step requests by interpreting customer-provided documents and images, applying policy, and using service tools, all in one reasoning loop
• Document intelligence — Interpret complex documents, including OCR, tables, charts, and page layouts, to extract the needed information and generate summaries or reports
• Enterprise productivity — Complete multi-step tasks by interpreting visual information, reasoning across it, and using tools, from screenshots and dashboards to GUIs the agent has to operate
• Video understanding — Extract key information from video to support search, content review, advertising, retail, and incident analysis
• Web research — Search across websites, connect information from multiple sources, and verify results, reading pages the way a person does

The common thread: these workflows used to require a vision model, an OCR pass, and a text model stitched together. With Nemotron 3.5 Super VL, perception and execution live in the same model, so context survives from the first page view to the final tool call.

Why run Nemotron 3.5 Super VL on FriendliAI?

Multimodal agents can be demanding to serve. Images and documents increase input size, multi-turn workflows reuse context, and a single task may involve multiple model calls.

FriendliAI Dedicated Endpoints provide a way to deploy Super VL on dedicated inference infrastructure.

FriendliAI’s is designed for demanding workloads, with optimizations for batching, throughput and long-context serving.

For developers, the goal is straightforward: evaluate Super VL on representative requests, measure latency and throughput, and configure the deployment around the application's needs.

What FriendliAI brings to Nemotron 3.5 Super VL:

Put Nemotron 3.5 Super VL into production

Nemotron 3.5 Super VL is available on Friendli Dedicated Endpoints starting today.

1️⃣ Navigate to the dedicated endpoint creation page.

2️⃣ Choose the model: [nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16]

3️⃣ Click "Create"

Prerequisites

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

Install the package

uv pip install friendli

Environment Setup

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

Example Code

Nemotron 3.5 Super VL accepts images through the standard OpenAI-compatible chat format, so you can send a document page or screenshot alongside your instruction:

import osfrom friendli import SyncFriendliwith SyncFriendli(    token=os.environ["FRIENDLI_API_KEY"],) as friendli:    res = friendli.dedicated.embeddings.embeddings(
        model="your-dedicated-endpoint-id",
        input_="Hello, Nemotron 3.5 Super VL!",
        encoding_format="float",
    )
# Handle responseprint(res.choices[0].message.content)

For teams with specialized tools, policies, or domain data, Nemotron 3.5 Super VL's open weights provide a clear path to post-train the model with NVIDIA NeMo for your own workflows, and deploy the customized version 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 3.5 Super VL gives agents a single model that can see, reason, and act across documents, images, and video, and FriendliAI is your platform to take it to production.

👉 Deploy Nemotron 3.5 Super VL on FriendliAI today and give your agents an execution layer that can see what they're working on.