NVIDIA's open-model giveaway: the Nemotron 3 era, mapped for builders

Quietly, the most prolific supplier of open-weight AI models in 2026 is not an AI lab — it's NVIDIA. The company gives away frontier-class models, the datasets used to train them, and the serving stack to run them, across language, speech, robotics, biology, and driving. The motive isn't charity: every downloaded checkpoint is future GPU demand. But for builders, motive doesn't matter — the resources do. Here's the map.

Nemotron 3: the flagship open family

Nemotron is NVIDIA's agent-focused open model line, and the Nemotron 3 generation rolled out across the past seven months on a hybrid latent mixture-of-experts architecture — big total parameter counts, small active counts, which is the whole efficiency trick:

model    total params   active/token   released
Nano     31.6B          ~3B            Dec 15, 2025
Super    ~120B          ~10B           Mar 11, 2026 (GTC)
Ultra    ~550B          ~50B           Jun 4, 2026 (Computex)

+ Nano Omni (Apr 2026): vision + audio + language unified
  in one open model, built for multimodal agents

Positioning is explicit: Nano for high-volume, targeted tasks (and it fits on a single consumer-class GPU quantized), Super for multi-agent applications, Ultra as the reasoning engine for coding assistants, search, and workflow automation. The family has passed 50 million downloads in a year — these aren't vanity releases, they're becoming the default self-hosted agent brain.

Notably, NVIDIA also releases much of the training and post-training data alongside the weights — the Nemotron datasets on Hugging Face are some of the largest openly published post-training corpora anywhere. If you're fine-tuning anything, that data is a resource independent of the models.

The rest of the open-model zoo

• Speech — Parakeet & Canary. Parakeet (FastConformer ASR; the current parakeet-tdt-0.6b-v3 covers 25 European languages with auto language detection) is the workhorse open transcription model — 600M params, runs anywhere. Canary does simultaneous transcription + translation across 25 languages. If you're adding voice to an app, these replace a paid ASR API outright.
• Physical AI — Cosmos 3. World foundation models: video-in, physics-consistent prediction out. Used to generate synthetic training data and give robots/AVs a world model.
• Robotics — Isaac GR00T N1.7. Open foundation model for humanoid and general robot control.
• Driving — Alpamayo; biomedical — BioNeMo/Clara — same pattern: open models + datasets for a vertical NVIDIA wants GPU-heavy.

Where everything lives

• huggingface.co/nvidia — the weights and datasets. This is the canonical open distribution point.
• build.nvidia.com — every model as a hosted NIM API you can hit with free credits before committing to self-hosting. Try Ultra here; download Nano there.
• OpenRouter / Azure AI Foundry — third-party hosted access if you want Nemotron behind a standard OpenAI-style endpoint.
• The serving stack is open too: NeMo (training + customization, now its own GitHub org), TensorRT-LLM (optimized inference), and Dynamo (distributed inference serving) are all Apache/permissive open source.

The license, honestly

Most of these ship under the NVIDIA Open Model License: commercial use allowed, derivatives allowed, no claim on your outputs. It is not OSI-approved open source — there are responsible-use terms — and some models ship under model-specific variants, so read the license file on the exact checkpoint before you build a business on it. In practice it's more permissive than Llama's license and less restrictive than most "open" model terms; that's a real, deliberate strategy to become the default.

Why this matters for agent (and crypto) builders

Our readers mostly build agents and on-chain apps, so the relevant lens: an agent that holds keys shouldn't leak its context to a third-party API. The moment your agent signs transactions — trading bots, wallet copilots, x402-paying services — every prompt you send to a hosted LLM is strategy, balances, and intent handed to someone else's logs. Open weights are the fix: Nemotron Nano on your own box is a competent agent brain with zero data egress, and the per-token cost of a self-hosted Nano is effectively electricity. The economics of agent-to-agent commerce only work when the marginal LLM call is near-free — which is exactly what NVIDIA is selling, one GPU at a time.

Resources

• Developer hub: developer.nvidia.com/topics/ai/nemotron
• Models & datasets: huggingface.co/nvidia
• Hosted APIs (free credits): build.nvidia.com
• GitHub: NVIDIA, NVIDIA-NeMo (the NeMo framework moved to its own org), TensorRT-LLM, and ai-dynamo/dynamo for inference serving.
• Announcements: Nemotron 3 launch and the research page.

TL;DR

• NVIDIA is 2026's biggest open-weights supplier: Nemotron 3 (Nano 31.6B / Super ~120B / Ultra ~550B hybrid MoE, + multimodal Nano Omni), Parakeet/Canary for speech, Cosmos for world models, GR00T for robots — plus training datasets and an open serving stack.
• Everything lives on huggingface.co/nvidia; try it hosted with free credits on build.nvidia.com.
• License is commercial-friendly but not OSI open source — read the checkpoint's license before shipping.
• For agents that hold keys, open weights aren't ideology — they're opsec and unit economics.