GGUF Discovery

Professional AI Model Repository

GGUF Discovery

Professional AI Model Repository

5,000+
Total Models
Daily
Updates
Back to Blog

NVIDIA Nemotron AI Models 2026: Nano 30B, Super 120B & Ultra 550B Complete Guide

Introduction to NVIDIA Nemotron

NVIDIA's Nemotron family has grown from Llama-derived enterprise models into the Nemotron 3 generation — a line of hybrid Mamba-Transformer Mixture-of-Experts (MoE) models with built-in reasoning (thinking mode) and native 1-million-token context windows. These are the first models where NVIDIA's silicon and software stack (NeMo, TensorRT-LLM) meet a genuinely open-weight frontier family, and they are among the most GPU-friendly models to deploy locally.

What sets Nemotron apart is its efficiency focus: even the flagship Ultra-550B activates only 55B parameters per token, while the Nano-30B activates just ~3B. That combination of huge total capacity with tiny active footprints makes Nemotron models unusually practical across the whole spectrum of consumer to datacenter hardware.

The Nemotron 3 Family at a Glance

Model Total Params Active Params Context Focus
Nemotron-3-Ultra-550B-A55B 550B 55B 1M tokens Flagship open-weight agentic
Nemotron-3-Super-120B-A12B 120B 12.1B 1M tokens Balanced reasoning workhorse
Nemotron-3-Nano-30B-A3B 31.6B 3.2-3.6B 1M tokens Local/edge reasoning
Nemotron-Cascade-2-30B-A3B 30B 3B 1M tokens Math/IMO-grade reasoning
Nemotron-Super-49B-v1 49B dense 128K Enterprise throughput
Nemotron-Hyper-245B (GenRM) ~245B ~22B 128-256K Reward modeling / alignment

Nemotron-3-Nano-30B-A3B: Local Reasoning Champion

Released December 2025, the Nano-30B is the star of the family for local deployment. Its hybrid architecture interleaves Mamba-2 state-space blocks with GQA attention and 128 routing experts (6 active per token), giving it ~3.3× higher throughput than standard 30B open models on a single H200.

Key Benchmark Results:

  • MMLU-Pro: 78.30 — outstanding for a 3B-active model
  • GPQA Diamond: 73.04 (75.00 with Python tool execution)
  • AIME 2025: 89.06 (99.17 with code sandbox tools)
  • LiveCodeBench v6: 68.25
  • IFBench: 71.51 — instruction following

A configurable reasoning/thinking-token budget lets you trade speed for depth, and quantized GGUF builds are available in the model repository — the ~3B active count means Q4 quantizations run on consumer hardware with 16-32GB of RAM.

Nemotron-3-Super-120B-A12B: The Reasoning Workhorse

Released March 2026, Super-120B uses a Hybrid Mamba-Transformer LatentMoE with NVFP4 precision and Multi-Token Prediction layers. Its 12.1B active parameters make it a strong middle ground between the Nano and Ultra tiers, evaluated rigorously on AIME 2025/2026, HMMT, Math 500, LiveCodeBench, and BFCL v4 function-calling benchmarks.

Nemotron-3-Ultra-550B-A55B: The Open-Weight Agentic Flagship

The June 2026 flagship is NVIDIA's most capable open-weight agentic model, built with LatentMoE and advanced multi-environment RL post-training:

  • GPQA Diamond: 86.7 — elite graduate-level science
  • τ²-Bench (Telecom/Retail/Airline): 83.3 — real-world agent execution
  • Coding Index: 49.3 — strong open-weight coding
  • Context: 1M tokens for repository-scale agent workflows

Nemotron-Cascade-2-30B-A3B: IMO-Grade Math

Also released March 2026, Cascade-2 post-trains the Nano-30B base via on-policy distillation and Cascade RL, achieving gold-medal-level performance on international math challenges (2025 IMO datasets) while closing the reasoning/knowledge gap on MMLU-Pro and GPQA.

Nemotron-Super-49B and the Enterprise Tier

The earlier Nemotron-Super-49B-v1 (128K context, dense) remains a popular high-throughput enterprise model for IT automation, summarization, and corporate workflows, while Nemotron-Hyper-245B (a Qwen-3-Nemotron-235B derivative) serves as NVIDIA's generalist reward model (GenRM) for RLHF alignment. The Super-49B has GGUF builds in the model repository, and the Super-120B (March 2026) extends the family with 12.1B active parameters and a 1M-token context for heavier reasoning workloads.

Running Nemotron Models Locally

  • 16-32GB RAM: Q4_K_M quantizations of Nemotron-3-Nano-30B-A3B and Cascade-2 — 1M context on consumer hardware
  • 32-48GB RAM: Nemotron-Super-49B Q4/Q5 builds, or Nano-30B with GPU offload
  • GPU (24GB+ VRAM): Super-120B Q4 partial offload for fast reasoning workloads

Conclusion

NVIDIA's Nemotron 3 family proves that datacenter-scale capability and local efficiency can coexist. From the Nano-30B's AIME 89% at just 3B active parameters to the Ultra-550B's GPQA 86.7%, Nemotron offers one of the most complete open-weight lineups — especially if you want reasoning, huge context, and GPU-native performance all in one family.