NVIDIA Nemotron for Real Estate Agents: A Free API With a Model for Every Job

All Things AI · Season 3 · The New Models · Episode 5

Nemotron is NVIDIA’s own family of open models, and it’s built so there’s a right-sized piece for every job: a fast one for everyday questions, a much larger one for the hardest reasoning, and dedicated pieces for reading a video, transcribing a call, or telling two speakers apart, all reachable through NVIDIA’s own free API.

What it is

Nemotron is a family of open models NVIDIA built and released with open weights, open training data, and open recipes, rather than one single model. That matters because it means you’re not stuck with one size doing every job.

The family has several parts: Lightning, a fast everyday model; Super and Ultra, larger models for harder reasoning; Nano Omni, which reads video, audio, and images together; plus dedicated speech-to-text and speaker-identification models.

Nemotron Lightning: the fast, everyday model

Lightning is a 30-billion-parameter model where only about 3 billion of those parameters are active on any given question, which is exactly why it’s fast. NVIDIA describes it as the fastest model of its size with leading accuracy for specialized agentic tasks.

It reads up to a million tokens of context, it can switch its own reasoning mode on or off depending on how hard the question is, and it works in several languages beyond English, including Spanish, French, German, Italian, and Japanese.

Nemotron Super and Ultra: for the harder problems

When a question needs more depth than Lightning gives you, Super and Ultra are the larger options in the same family. Super is a 120-billion-parameter model NVIDIA points at agentic workflows, long-context reasoning, and high-volume tasks like sorting a large batch of tickets. Ultra is the largest at 550 billion parameters, aimed at the hardest reasoning and the most complex multi-step agent work.

Both read up to a million tokens of context, the same as Lightning, so the difference between the three is depth of reasoning rather than how much you can hand them at once.

Nemotron Nano Omni: reading video, audio, and documents together

Nano Omni takes video, audio, images, and text all at once and answers in text. NVIDIA points it at jobs like reviewing a meeting recording, reading a complex business document with charts in it, or handling a task that needs several kinds of input read together.

There are also two dedicated pieces worth knowing: a speech-to-text model that transcribes with proper punctuation and capitalization across 40 languages, and a diarization model that answers the specific question of who spoke when in a recording with up to eight different speakers.

A safety layer built for the same family

Alongside the models that actually answer questions, NVIDIA also released a compact guardrail model as part of the same family, built specifically to check both what goes into a larger model and what comes back out of it. It’s a small model on purpose, at only about 4 billion parameters, which keeps that checking step fast rather than adding a slow extra hop before every answer.

For anyone building a workflow where an AI model is handling something client-facing, that kind of checking layer is a genuinely useful piece to know exists, even if you never touch it directly yourself.

The free API: where all of it lives

This is the part worth spending the most time on. build.nvidia.com is NVIDIA’s own catalog where you can try any of these models directly in a browser or reach them through an API, and members of NVIDIA’s Developer Program get free access to prototype with them.

This is a topic Victoria has said she loves talking about, because it means no separate signup with each individual model maker, one free account, and the whole family of models available from the same place.

  • Sign up once at build.nvidia.com and every model in the family is reachable from the same account.
  • Try any model directly in the browser before deciding whether to build it into a workflow.
  • The free tier is meant for prototyping and everyday use; a separate production path exists for anyone running it at real scale.

What to use it for

The whole point of a family this size is picking the right-sized model for the job instead of paying for more model than the question needs.

  • Quick, everyday questions and drafts on Lightning, where speed matters more than depth.
  • A genuinely hard reasoning problem, a long document, or a complex multi-step task on Super or Ultra.
  • Reading a meeting recording or a document full of charts on Nano Omni in one pass.
  • Getting a clean, punctuated transcript of a call or recording through the dedicated speech model.
  • Figuring out who said what in a recording with several speakers through the diarization model.

Where it is weaker

Two honest limits.

The larger models, Super and especially Ultra, are built for depth rather than speed, so a quick, simple question doesn’t need them. Lightning already covers most everyday work.

And running any of these on your own hardware, rather than through the free API, takes serious equipment for the bigger models, real datacenter-class GPUs. For almost every agent, the free hosted door on build.nvidia.com is the practical way in, not self-hosting.

Why an open family, specifically, is worth paying attention to

NVIDIA’s own description of Nemotron leans hard on the word open: open weights, open training data, and open recipes for how the models were built, not just a downloadable file with no explanation behind it. That openness is also what makes it possible to run smaller pieces of the family, like Lightning, on local devices such as an RTX 5090 or an NVIDIA DGX Spark, rather than only through a hosted service.

It also means the family keeps growing in public. NVIDIA has continued releasing new pieces on a regular cadence throughout 2026, most recently the dedicated diarization model, which means checking back periodically is worth doing rather than treating this as a one-time announcement.

What to try in the first week

Two exercises, both aimed at feeling the difference the free API makes.

Sign up once at build.nvidia.com, and run the same real question through Lightning first, then through Super or Ultra, and compare the depth of the two answers side by side.

Then take a recording of one of your own team meetings with more than one speaker, and try the diarization model on it, asking who spoke when. That’s the specific job most everyday tools don’t handle well.

How it sits next to the others

Other models in this season are one size doing every job. Nemotron’s whole design is the opposite: a right-sized piece for each kind of question, all under one free account.

For an agent building a habit of trying different tools, this is the one place where you can compare a fast model against a much deeper one, or try a dedicated transcription and speaker-identification tool, without opening five separate accounts to do it.

Questions agents actually type

Is NVIDIA Nemotron free to use?

Yes, through build.nvidia.com. Members of NVIDIA’s Developer Program get free access to prototype with the whole family of Nemotron models from one account, both in the browser and through an API.

What is the difference between Nemotron Lightning, Super, and Ultra?

They’re the same family at three different depths. Lightning is the fastest, built for everyday questions. Super and Ultra are larger models for harder reasoning and more complex multi-step work, with Ultra the largest and deepest of the three.

Can Nemotron read a video or a meeting recording?

Yes, through Nemotron Nano Omni, which reads video, audio, images, and text together and answers in text. It’s built for jobs like reviewing a meeting recording or a document full of charts.

Can Nemotron tell me who said what in a recording?

Yes, through its dedicated diarization model, which is built specifically to answer who spoke when in a recording with up to eight different speakers.

Do I need to run Nemotron on my own computer?

No. The practical way in for almost every agent is the free hosted API at build.nvidia.com. Running the larger models yourself takes serious, datacenter-class equipment, though the smaller Lightning model can run on a strong local device.

Book a 30-minute strategy call →

Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.