Nemotron 3 Super
NVIDIA
Key facts · verified Oct 29, 2026
Nemotron 3 Super is a NVIDIA model released on 11 Mar 2026. Pricing: Open weights — free to self-host. Its context window is 1M tokens (256k default, VRAM-limited). It ranks #210 on LMArena's text leaderboard (Elo 1360, read 1 Nov 2026). Source: developer.nvidia.com.
A 120B hybrid Mamba-Transformer MoE (12B active) under the open NVIDIA Nemotron Open Model License, sitting between Nemotron 3.5 Lightning and Nemotron 3 Ultra in this lineup. Tuned for agentic throughput: NVIDIA reports 5x higher throughput than the prior Nemotron Super.
Pricing & specs checkedSource: vendor documentation · checked Oct 29, 2026LMArena: 1360 Elo (rank 210) · source · retrieved Nov 1, 2026
Pricing
Open weights — free to self-host
Context
1M tokens (256k default, VRAM-limited)
LMArena Elo
LMArena rank
Key Features
Open Weights Hybrid Mamba-MoE Agentic Throughput Long Context Self-Hostable
Full Specifications
Open weights — free to self-host | |
|---|---|
| Identity | |
| Developer | NVIDIA |
| Released | Mar 2026 |
| Status | GA |
| Licence | Open weights (NVIDIA Nemotron Open Model License) |
| Self-hostable | Yes |
| Cost | |
| Cached input | Not published |
| Batch discount | Not applicable |
| Free tier | Free on build.nvidia.com and some third-party hosts |
| Capacity | |
| Context window | 1M tokens (256k default, VRAM-limited) |
| Max output | Not published |
| Long-context surcharge | Not applicable |
| Capability | |
| Vision in | No |
| Audio in | No |
| Function calling | Yes |
| Structured output | Not published |
| Extended reasoning | Not published |
| Web search | No |
| Code execution | No |
| Access | |
| API | Via build.nvidia.com and third-party hosts |
| Chat app | No (open weights) |
| Cloud marketplaces | Google Cloud Vertex AI, Oracle Cloud, CoreWeave, Together AI |
| Fine-tuning | Yes (open weights) |
| Measured quality | |
| LMArena Elo | 1360 (checked Nov 2026) |
| LMArena rank | Rank 210 |