Nemotron models
Only Nemotron is supported as a first-party local model. That is a deliberate product decision: one model family means one set of provider profiles, one set of policy profiles, and one thing to test.
| Model | Parameters | Context | Notes |
|---|---|---|---|
| Nemotron Nano 12B v2 VL | 12.6B dense | 128K | vision: up to 5 images, documents, video |
| Nemotron 3 Nano | 30B-A3B (3.2B active) | 1M | fastest in the family, hybrid Mamba-Transformer MoE |
| Nemotron 3 Super | 120B-A12B | 1M | multi-agent and coding; 85.6 % PinchBench (OpenClaw) |
| Nemotron 3 Ultra | 550B-A55B | 1M | strongest open model in the family, NVFP4 |
Sources and dates are in ../claims.md.
How a nest reaches a model
Nestbox never injects a model endpoint into an agent's code. It registers a provider profile and attaches it to the nest:
- Cloud — a NIM endpoint on
build.nvidia.com, or any OpenAI-compatible provider. - Local — a
local-nemotronnest running vLLM, reachable over the overlay athttp://<overlay-ip>:8000/v1. - Your laptop — the same thing, on your laptop, over WireGuard.
The policy does not change between these. What changes is one profile.
VRAM guidance
See ../hardware/ for the sizing table and the caveats. The short version:
- 12B VL → 24–32 GB VRAM
- 3 Nano → 32–48 GB
- 3 Super → 80 GB
- 3 Ultra → 4× 96 GB or a DGX B200
Context length is the hidden cost: a 1M-token window needs hundreds of GB of KV cache, so cap the context on single-GPU deployments (32K–128K is the practical default) unless the hardware is sized for more.