LLM
GA
Open weights
Nemotron 3 Ultra
Synthetic data generation, distillation, and large-scale open reasoning
Release2026-03
Context1M
Max output—
Parameters550B
Local / VRAM640–1000 GB
LicenseOpen weights
Overview
Nemotron 3 Ultra is a model in the LLM category from NVIDIA. It was released in 2026-03. The vendor is based in US. Best for: Synthetic data generation, distillation, and large-scale open reasoning.
Hardware: FP8 ~550 GB; DGX/HGX B200 class.
Strengths
- 1M context
- Open weights
- NVIDIA NIM optimizations
Limits
- Very large hardware
- Commercial license but conditional
Specifications
| Vendor | NVIDIA |
|---|---|
| Category | LLM |
| Release | 2026-03 |
| Status | GA |
| License | Open weights |
| License name | NVIDIA Open Model License |
| Commercial use | Allowed |
| Parameters | 550B |
| Active parameters | 55B |
| Architecture | MoE transformer |
| Context window | 1M |
| Input modalities | Text, Image |
| Output modalities | Text |
| Prompt types | Text |
| Negative prompts | No |
| Max outputs | 1 |
| Local deployment | Yes |
| VRAM | 640–1000 GB |
| System RAM | 768–1536 GB |
| API access | Yes |
| Pricing | Free weights / NIM API |
| Confidence | Medium |
| Vendor country | US |
| Vendor type | Open lab |
Pricing & hardware
Free weights / NIM API
This model can be self-hosted or run locally. Approximate GPU memory: 640–1000 GB. Approximate system RAM: 768–1536 GB. API access is available.