n8n shows how to run AI workflows on one NVIDIA DGX Spark in the office, with no cloud API
A new guide on the n8n blog is aimed at small businesses that cannot send customer data to a cloud API, such as clinics, law firms and accountants. The post shows how to run every model that n8n workflows need on one NVIDIA DGX Spark with 128GB of shared memory: one MoE model, Qwen3.6-35B-A3B, for chat, data extraction and reading screenshots, alongside models for embeddings, reranking, transcription, OCR, image generation and text to speech. All the models sit behind one OpenAI-compatible LiteLLM endpoint with a separate key per team, and n8n connects to it through a regular OpenAI credential with a custom Base URL. The post includes a workflow template that reads an invoice uploaded to a form and returns the supplier, invoice number, date and total with no cloud call at all. With the model's thinking step turned off, extraction dropped from 12.1 seconds to 0.59 seconds. In the author's measurements, the box used 108GB with seven models loaded, and chat slowed to about a third of its speed while images were being generated. The post recommends choosing only models whose license allows commercial use.
Otomatzia editorial · Published
Source published: · n8n
