Nemotron 3 Nano vs Super vs Ultra: Which NVIDIA Model?
The short answer
Nano for cheap, fast agent steps; Super for multi-agent coding on one 8-GPU node; Ultra for maximum accuracy. All three are NVIDIA’s open Nemotron 3 models — not Meta’s — sharing one design: a hybrid Mamba-Transformer mixture-of-experts with up to 1 million tokens of context and reasoning-budget control at inference time. They differ in size by roughly 4x at each step, and Super and Ultra add LatentMoE, multi-token prediction layers and NVFP4 pretraining. Facts below are as of October 15, 2026.
The comparison
| Nemotron 3 Nano | Nemotron 3 Super | Nemotron 3 Ultra | |
|---|---|---|---|
| Released | December 15, 2025 | March 2026 | June 4, 2026 |
| Total / active parameters | 31.6B / 3.2B (3.6B with embeddings) | 120B / 12B | 550B / 55B |
| Architecture | Hybrid Mamba-Transformer MoE | + LatentMoE, MTP layers, NVFP4 pretraining | + LatentMoE, MTP layers, NVFP4 pretraining |
| Context | Up to 1M tokens | Up to 1M tokens | Up to 1M tokens |
| Minimum hardware | One data-center GPU | 8x H100-80GB (NIM docs) | Multi-GPU server; NVFP4 checkpoint for Blackwell |
| NVIDIA’s positioning | Cost-efficient accuracy | Collaborative agents, high-volume work, coding, cyber triage | Frontier reasoning for long-running agents |
| Open release | Weights, recipe, data | Weights, data, recipes | Weights (BF16, NVFP4, base), data, recipes |
How they differ in practice
Nano: the throughput model. With 3.2B active parameters, Nano is cheap to run at volume. NVIDIA reports 3.3x the throughput of Qwen3-30B-A3B and 2.2x that of GPT-OSS-20B on a single H200 (8K input, 16K output), and higher accuracy than both on its benchmark set. Use it for routing, extraction, tool-call steps and local agents. Note that Tinker lists Nemotron 3 Nano as retiring from its training platform on October 23, 2026, alongside the newer Nemotron 3.5 Lightning at the same 30B-A3B size.
Super: the multi-agent workhorse. NVIDIA built Super for the “context explosion” of multi-agent systems, which it says generate up to 15x the tokens of a standard chat. It claims over 5x the throughput of the previous Nemotron Super and scores 85.6% on PinchBench, a benchmark of models acting as the brain of an OpenClaw agent. Super supports English, French, German, Italian, Japanese, Spanish and Chinese and is cleared for commercial use. Choose it when one 8-GPU node is your budget and Nano’s accuracy is not enough.
Ultra: the accuracy pick. Ultra is the final and largest Nemotron 3 model. NVIDIA reports 5.9x the inference throughput of GLM-5.1 (754B-A40B), 4.8x that of Kimi K2.6 and 1.6x that of Qwen 3.5-397B at 8K input and 64K output, with on-par accuracy against leading open models and better RULER scores at 1M tokens. It is available on build.nvidia.com, Hugging Face, Amazon SageMaker JumpStart and OpenRouter, so you can test it before buying hardware.
Which one to pick
- Agent steps at scale, local or edge-adjacent: Nano, or Nemotron 3.5 Lightning for new projects.
- Coding agents and multi-agent pipelines on one node: Super.
- Hardest reasoning, long autonomous runs, hosted access is fine: Ultra.
- Fine-tuning on a budget: all three are on managed trainers billed by active parameters — see fine-tuning platforms with token billing.
Open weights do not mean free to run: price the GPUs against hosted open models in our current API prices before you self-host.
Last verified: October 15, 2026.