Claims and sources
Every performance or capability claim in this repository must appear here with a source and a date. Marketing copy cites this file rather than repeating numbers.
| Claim | Source | Date |
|---|---|---|
| Nemotron 3 Nano: 30B total, 3.2B active, hybrid Mamba-Transformer MoE, 1M context | research.nvidia.com/labs/nemotron/Nemotron-3/ | 2025-12-15 |
| Nemotron 3 Super: 120B total, 12B active, NVFP4, 1M context, 5× throughput vs previous Super | developer.nvidia.com blog, "Introducing Nemotron 3 Super" | 2026-03-11 |
| Nemotron 3 Super: 85.6 % on PinchBench for OpenClaw agents | same blog post | 2026-03-11 |
| Nemotron 3 Ultra: 550B total, 55B active, NVFP4, 1M context, on-par accuracy with state-of-the-art open LLMs, best RULER at 1M | research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/ | 2026-06-04 |
| Nemotron 3 Ultra: 5.9× / 4.8× / 1.6× throughput vs GLM-5.1-754B-A40B / Kimi-K2.6-1T-A32B / Qwen-3.5-397B-17B | same page | 2026-06-04 |
| Nemotron Nano 12B v2 VL: 12.6B, multi-image (up to 5), video, 128K context | build.nvidia.com/nvidia/nemotron-nano-12b-v2-vl | 2025-10-28 |
| Nemotron 4 is in development | NVIDIA GTC 2026 keynote | 2026 |
Rules
- A claim without a row here does not ship.
- Comparisons against proprietary models are quoted, with a timestamp and a video link, never asserted in our own voice.
- VRAM figures are labelled as our estimates, because the vendor publishes none.
- Trademarked product names of other vendors are not used in comparative claims.