Nemotron Nano 3 30B-A3B

Overview

NVIDIA's Mamba-2 + Transformer hybrid MoE with 3B active out of 30B total parameters. A native 1M-token context with roughly 4× the throughput of Nemotron 2.

When to pick this model

• Million-token context workloads
• Edge and on-device inference at unusually long context
• Throughput-critical pipelines (RAG ingestion, log analysis)
• Hybrid SSM-Transformer research and benchmarking

VRAM requirements by quantization

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

Strengths

• Native 1M-token context window
• Ultra-efficient MoE with only 3B active parameters
• Roughly 4× throughput improvement over Nemotron 2
• Permissive NVIDIA Open Model license

Limitations

• Full 1M context consumes substantial VRAM in practice
• Hybrid architecture has thinner tooling support
• Distilled from Llama — inherits some base-model quirks

Architecture & training

Architecture: MoE · 30B total / 3B active · Nemotron-Nano-3 · 1M native context

Training: NVIDIA — distilled from Llama, edge-optimized with 1 million token context.

Quick start

ollama run nemotron3:30b

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.