Preview NVIDIA

NVIDIA VSS blueprint / Cosmos Reason

NVIDIA's Video Search and Summarization blueprint runs VLMs (Cosmos Reason, Qwen3-VL) over live RTSP streams for captions and alerts. It is self-hosted; build.nvidia.com offers free preview API calls, not a hosted streaming endpoint.

Est. per minuten/a

Overview

Best for: Enterprises with GPUs building camera monitoring and video search.

At a glance

Free tierYes

CPU onlyNo

WebRTCNo

Self-hostYes

Open weightsNo

Self-hosted blueprint over RTSP on your own GPUs; production needs NVIDIA AI Enterprise licence. Free preview API credits (up to 5,000) for trials only. Chunk-based, not frame-level.

Audio in

n/a

Audio out

n/a

Languages

n/a

Latency

Chunk-based (user-defined chunk duration); not frame-level realtime.

Regions

Wherever you deploy.

Compliance

Depends on your deployment.

Features

live RTSP stream alerts
captioning and anomaly detection
self-hosted NIM microservices

Pricing

What	Price	Unit	Notes
build.nvidia.com preview APIs	Up to 5,000 free credits	per new account	Trial only
Production self-hosting	NVIDIA AI Enterprise licence		Plus your own GPUs

How the per-minute estimate was worked out

Self-hosted; depends on your GPUs and licence

Free tier: Preview API credits for new accounts.

Source: nvidia.com

Setup

Deploy the VSS blueprint on your GPUs (see docs.nvidia.com/vss).
Register RTSP streams and configure chunk duration and alert prompts.
Use build.nvidia.com preview APIs only for trials.

Endpoint

Self-hosted VSS REST API

Authentication

NGC / NVIDIA API key

Warnings

Not a hosted realtime API

Plan on running GPUs yourself; the hosted catalog is for trials.

Licence needed for production

Downloadable NIMs need an NVIDIA AI Enterprise licence in production.

Chunked, not per-frame

Alerts run on segments at an interval you set, so reaction time is at least one chunk.

Plus 9 warnings that apply to all avatars and live video APIs. See category warnings.

Limits

No hosted realtime streaming endpoint confirmed

Models and products

Name	Status	Notes
cosmos-reason2-8b	GA	Downloadable NIM; 256K input context.
VSS 3.2 Real-Time VLM component	GA	Self-hosted; processes stream chunks at a user-defined interval.

Docs and sources

Docs

Sources used

Not fully verified

Whether any hosted streaming endpoint exists; Cosmos 3 status.

Similar avatars + video APIs

SynthesiaSynthesia Interactive Avatarsn/a1 high-severity warnings OdysseyOdyssey API (interactive world model video)n/a1 high-severity warnings Daydream / LivepeerDaydream (hosted StreamDiffusion)n/a1 high-severity warnings Google DeepMindGoogle Project Genie (Genie 3)n/a1 high-severity warnings

Spotted a wrong price or a dead link?