NVIDIA announced the Nemotron 3 family at GTC 2026 — a set of open models designed specifically for NVIDIA hardware. They range from a 4B model that runs on a Jetson to a 120B model for datacenter GPUs. The twist: NemoClaw, an agent framework that turns these models into persistent AI workers.

The family at a glance

Model	Parameters	Target hardware	Context	License	Best for
Nemotron 3 Nano 4B	4B	Jetson, RTX laptops	32K	Open	On-device, edge AI
Nemotron 3 8B	8B	RTX 3060+	64K	Open	Desktop AI assistant
Nemotron 3 Super 120B	120B (MoE)	A100/H100	128K	Open	Enterprise, datacenter
NemoClaw	Framework	Any NVIDIA GPU	—	Open	AI agent deployment

All models are open source and optimized for NVIDIA’s TensorRT-LLM inference engine.

Nemotron 3 Nano 4B

The smallest model, designed for NVIDIA Jetson devices and RTX laptops. At 4B parameters, it fits in 3 GB of VRAM and runs at 50+ tokens per second on an RTX 4060.

What makes it special

NVIDIA trained Nano specifically for on-device use cases: voice assistants, real-time translation, local document processing. It’s not trying to compete with Gemma 4 or Llama 4 on benchmarks — it’s optimized for speed and efficiency on NVIDIA silicon.

Setup

# Via Ollama
ollama run nemotron3:nano

# Via TensorRT-LLM (fastest on NVIDIA GPUs)
pip install tensorrt-llm
trtllm-build --model nemotron-3-nano-4b --output ./engine

The TensorRT-LLM path is 2-3x faster than standard inference on NVIDIA GPUs. If you have an RTX card, it’s worth the extra setup.

Hardware requirements

Quantization	VRAM	Speed (RTX 4060)
FP16	8 GB	40 tok/s
INT8	4 GB	55 tok/s
INT4	3 GB	65 tok/s

Compare this to Gemma 4 E4B which needs similar hardware but doesn’t have NVIDIA-specific optimizations. On NVIDIA GPUs, Nemotron Nano is faster.

Nemotron 3 8B

The mid-range model for desktop use. Comparable to Qwen 3.5 Flash and Llama 4 Scout 8B in quality, but with NVIDIA-specific optimizations.

ollama run nemotron3:8b

Benchmarks vs peers

Benchmark	Nemotron 3 8B	Gemma 4 E4B	Qwen 3.5 Flash
MMLU	72.1	70.8	71.5
HumanEval	68.3	65.2	70.1
GSM8K	78.5	76.2	77.8

Competitive but not dominant. The advantage is speed on NVIDIA hardware — TensorRT-LLM makes it 2-3x faster than running the same-size models through standard Ollama or llama.cpp.

Nemotron 3 Super 120B

The datacenter model. 120B parameters with MoE architecture, designed for A100 and H100 GPUs. This competes with Qwen 3.5 Plus and Llama 4 Maverick.

Hardware requirements

Setup	GPUs needed	Inference speed
FP16	2x A100 80GB	30 tok/s
INT8	1x A100 80GB	45 tok/s
INT4	1x A100 40GB	55 tok/s

This isn’t a laptop model. It’s for teams running self-hosted AI infrastructure. The advantage over competitors: NVIDIA’s inference stack is the most mature and best-supported in enterprise environments.

NemoClaw: AI agents on NVIDIA hardware

NemoClaw is NVIDIA’s version of OpenClaw — an agent framework optimized for NVIDIA GPUs. It lets you deploy persistent AI agents that run locally on your hardware.

Key differences from OpenClaw

	NemoClaw	OpenClaw
Hardware	NVIDIA GPUs only	Any (CPU, AMD, NVIDIA)
Inference	TensorRT-LLM	Standard Python
Speed	3-5x faster	Baseline
Models	Nemotron optimized	Any model
Setup	More complex	Simpler

Setup

# Requires NVIDIA GPU + CUDA
pip install nemoclaw

# Initialize with Nemotron 3 8B
nemoclaw init --model nemotron3:8b --gpu 0

# Run an agent
nemoclaw run my-agent

When to use NemoClaw vs OpenClaw

Use NemoClaw if: You have NVIDIA GPUs and need maximum inference speed. Enterprise deployments where agents run 24/7.

Use OpenClaw if: You want hardware flexibility, simpler setup, or need to run on CPU/AMD GPUs.

Who should care about Nemotron 3?

NVIDIA GPU owners

If you have an RTX 3060 or better, Nemotron 3 models with TensorRT-LLM are the fastest local AI option available. The speed difference is significant — 2-3x over standard inference.

Enterprise teams

The combination of Nemotron 3 Super + NemoClaw gives enterprises a fully self-hosted AI agent stack with NVIDIA’s enterprise support. No cloud dependency, no data leaving your infrastructure.

Edge/IoT developers

Nemotron 3 Nano on Jetson devices opens up AI for robotics, industrial automation, and smart devices. It’s one of the few models specifically optimized for NVIDIA’s edge hardware.

How it compares to the competition

For most developers, Gemma 4 is a better starting point — it’s more portable, has a larger community, and runs well on any hardware. Nemotron 3 is the pick when you’re committed to NVIDIA hardware and want to squeeze maximum performance from it.

For a broader view of what’s available for local AI, see our best local AI models by task ranking and best GPU for AI locally guide.

If you’re deciding between running AI locally or using cloud APIs, our self-hosted AI vs API comparison covers the full tradeoff analysis.

FAQ

Is Nemotron free?

Yes, Nemotron 3 models are released under NVIDIA’s open model license which permits free commercial use. You can download and deploy them without licensing fees.

Which Nemotron model should I use?

Use Nemotron 3 8B for local development and lightweight tasks, and Nemotron 3 Super 120B for production workloads requiring frontier-level quality. NemoClaw is the right choice if you’re building AI agents on NVIDIA infrastructure.

Can I run Nemotron locally?

The 8B model runs easily on consumer GPUs with 8-12GB VRAM. The 120B Super model requires enterprise hardware — typically 2-4x A100 GPUs — but is still more practical to self-host than many competing models at that scale.

How does Nemotron compare to Llama?

Nemotron 3 Super 120B outperforms Llama 3.1 405B on several benchmarks while being significantly smaller and cheaper to run. The 8B variant is competitive with Llama 3.1 8B, with particular strengths in instruction following and tool use.

Related: Best AI Engineering Courses