Use NVIDIA Models in Big-AGI.

Frontier open models, for $0.

Every model below runs at no cost on NVIDIA's hosted catalog: NVIDIA's own Nemotron 3, plus GPT-OSS, DeepSeek V4, GLM, MiniMax, Mistral, Gemma, and Llama. One free nvapi- key, no card, no credits to burn down. The throttle is on requests, not tokens, so long prompts and long answers go through where token-capped free tiers stop you.

The limits are real and worth knowing: NVIDIA does not publish a rate, developers see somewhere around 40 requests per minute, and it moves with load. NVIDIA's trial terms also reserve the right to use your prompts and the answers to improve its products and models, and do not cover production use. Keep sensitive and production work on a paid provider.

Running NVIDIA in Big-AGI

Create an nvapi- key at build.nvidia.com, keep the Public API Endpoints scope on it, and paste it into Big-AGI. Nothing is billed for these models, and Big-AGI adds no markup and no intermediary: the key, the account, and the limits are yours, direct from NVIDIA.

• Rate limits, handled. The free endpoint throttles hard and unpredictably, so Big-AGI paces its calls and a Beam scatter queues instead of collecting 429s.
• A catalog that answers. NVIDIA reserves the right to drop any model at any time, and its model list is a stale superset: roughly half the ids it returns no longer answer. Big-AGI ships a curated list refreshed against the live endpoint, so a dead id never reaches your model picker.
• Context windows measured, not copied. The catalog pages misstate context length on roughly a quarter of these models, by as much as 8x, so Big-AGI probes the real limit against the endpoint.
• Reasoning controls wired through. GPT-OSS keeps its native low, medium, and high effort setting; the Nemotron thinking models get a thinking toggle.

Why Big-AGI instead of the catalog playground?

build.nvidia.com is a per-model demo box: one prompt, one model, no history. Big-AGI turns the same key into a workspace with persistent chats, personas, and attachments, and puts a free NVIDIA model in the same conversation as Claude, GPT, or Gemini, with the reasoning dials the demo box hides.

Your keys and your data

The default NVIDIA endpoint refuses browser-origin requests, so Direct Connection is off there and calls route through the Big-AGI server. Point the same service at your own host, a self-hosted NIM or vLLM box such as a DGX Spark serving an OpenAI-compatible port on your LAN, and Direct Connection works. Either way your keys stay in your browser, chats are stored locally first and sync only if you turn it on, and the AI Inspector shows the exact request.

NVIDIA in Beam

A free catalog earns its keep in Beam: run Nemotron 3 Ultra, GPT-OSS 120B, and DeepSeek V4 on one prompt at no cost, then reach for Fusions: several strategies that combine, cross-check, and synthesize the parallel answers, which beats just picking the best one. Parallel runs use more tokens than a single chat, and here they also spend more of that per-minute budget.