Automation
6 min read

Local AI with n8n: What a DGX Spark Handles in Testing, What It Costs and Where It Slows Down

Colorful cube of toy bricks with brick paths leading to six modules, surrounded by a low wall
Image: AI-generated
In short

According to an n8n test published on October 19, 2026, an NVIDIA DGX Spark with 128 GB can serve seven AI models to n8n at once and still deliver 47 tokens per second with eight users. Since October 2, though, the device costs $6,950, the new 64 GB version can't hold the full stack, and image generation slows chat down by roughly two thirds.

On October 19, 2026, the n8n team published a post on the n8n blog describing a complete AI stack that runs on a single NVIDIA DGX Spark and that n8n talks to just like a cloud API. It reads more like a lab report than marketing: it measures speed and memory and lists the places where the test hit snags. n8n names clinics, law firms, and accounting and tax offices as the target audience, meaning anyone who isn't allowed to send client data to cloud providers.

What runs on a single box in the n8n test?

Seven open models, all licensed for commercial use (Apache 2.0 or MIT):

  • Chat, vision and data extraction: Qwen3.6-35B-A3B. The model has 35 billion parameters, but only about 3 billion are active for each generated token. That makes it fast on this hardware.
  • Search across your own documents: Qwen3-Embedding-4B (turns text into numeric vectors so similar content can be found) and Qwen3-Reranker-0.6B (sorts the hits by relevance).
  • Speech recognition: Whisper large-v3-turbo.
  • Text recognition from images (OCR): GLM-OCR.
  • Images and text to speech: Z-Image-Turbo and Kokoro-82M.

The models run in vLLM, software that serves language models over an API. In front of it sits LiteLLM, a proxy that bundles all models behind one OpenAI-compatible endpoint and gives each team its own key. In n8n, you enter that endpoint as a custom Base URL in an OpenAI credential and then use the regular OpenAI nodes. Together, all seven models take up 108 of the 121 GB the system reports as available.

How fast is it in practice?

Fast enough for typical office workflows. Measured on October 13, 2026 with the chat model (a token is roughly a word fragment):

Concurrent users Time to first token Tokens per second per user
1 0.11 s 120
4 0.21 s 68
8 0.29 s 47

More numbers from the test: OCR on one page takes 1.80 seconds, and a 30-second audio clip is transcribed in 0.58 seconds. Structured data extraction takes 0.59 seconds, or 12.1 seconds with thinking mode turned on.

The big catch: when image generation runs at the same time, chat collapses. With one user, speed dropped from 122 to 38 tokens per second; with four users, from 69 to 24. That is 65 to 69 percent slower. The author recommends queuing image jobs or running them outside business hours.

What does the hardware cost since October?

A lot more than at launch. Nvidia raised the price of the 128 GB version to $6,950 on October 2, 2026, The Register reports. At launch it sold for around $4,000. The Register attributes the jump to sharply higher memory prices. Nvidia also announced a 64 GB version for around $4,999, sold only through partners such as Acer, Asus, Dell, Gigabyte, HP and MSI; the NVIDIA product page still lists it as "Coming Soon." Prices in euros vary by retailer and configuration, so compare current offers before you buy. The power supply is rated at 240 watts.

Our take: the 64 GB version is not enough for the stack in the test. According to the measurements, the chat model alone needs 43 GB, plus 11 GB for OCR, 8 GB for Whisper and 8 GB for the reranker. That already adds up to 70 GB. With 64 GB, you'll have to stick to the chat model plus one or two smaller services, or pick smaller models.

Which pitfalls should you know about first?

Here are the key lessons from the test as a checklist:

  1. Turn off the Responses API: From node version 1.3, the OpenAI Chat Model node in n8n uses the Responses API by default. The local models run on the Chat Completions format, so disable the "Use Responses API" option.
  2. Disable thinking mode for extraction: Otherwise an invoice record takes about twenty times as long (12.1 instead of 0.59 seconds).
  3. Validate OCR output: In the test, text recognition dropped a zero from a German VAT ID (USt-IdNr.). IDs, IBANs and amounts need a validation rule in your workflow. The form in the sample workflow only accepts PNG and JPEG, so you have to convert PDFs first.
  4. Trim silence from audio files: In the test, Whisper invented a short phrase at a silent ending that nobody had said.
  5. Keep LiteLLM up to date: The author requires at least version 1.84.0, pointing to two backdoored releases on PyPI in March 2026 and to authentication flaws.
  6. Check the license at the exact source: Some models have a non-commercial twin repository. The exact repository ID determines which license applies.
  7. Start models one after another: All services share one memory pool. Start them at the same time and you get out-of-memory errors. In the test, a cold start of the services took about twelve minutes.

When does a local LLM pay off for n8n workflows?

When sensitive data has to stay in-house and several people use AI workflows every day. For a first test, less is often enough:

Option Fits when Hardware Catch
Cloud API (OpenAI, Anthropic, Google) Data may leave under a data processing agreement, few runs none Data leaves your company, cost per use
n8n's Self-hosted AI Starter Kit (n8n, Ollama, Qdrant, PostgreSQL) Proof of concept on an existing machine existing Per n8n not optimized for production, no GPU inside the Docker container on a Mac
DGX Spark with vLLM and LiteLLM Sensitive data, one team, several AI services $6,950 (128 GB) Maintenance, memory ceiling, image generation slows chat

How Make and n8n compare on licensing and hosting is covered in our Make vs n8n cost comparison. If you still run an older n8n version, plan your n8n 3.0 migration first.

What does this mean for you?

  • Workflow first, hardware second. Build the process with test data and a cloud model or the Starter Kit. Count how many runs you get per day. Only then is it worth asking about your own device.
  • Test quality yourself. The n8n post measures speed and memory, not how good the answers are. Whether a model with 3 billion active parameters reliably reads your invoices, contracts or emails only shows in a test with 20 to 50 real examples.
  • Local does not automatically mean GDPR compliant. You still need access controls, logging and deletion periods. None of this replaces a review by your data protection officer.
  • One box, one point of failure. If the device goes down, every AI step stops. Build an error path in n8n that parks or reports tasks instead of silently losing them.

Our take: for law firms, medical practices, and accounting and tax offices with fixed processes such as receipt capture or dictation, this stack is now within reach, because n8n documents the stumbling blocks openly. For teams without in-house IT, the real cost is maintaining vLLM, LiteLLM and seven models, not the price of the device. We ran the numbers on self-hosting a German model using Aleph Alpha Kolibri as an example. If you would rather not set up the stack yourself, our n8n automation agency page shows how we work with n8n. If you want to check whether a local LLM makes sense for one of your processes, get in touch.

Frequently asked questions

Can n8n use local AI models without extra software?

Not directly. The models need a server that serves them, such as Ollama in the Starter Kit or vLLM with LiteLLM as in the DGX Spark test. n8n then connects to that server through an OpenAI credential with a custom Base URL.

Is a regular office PC enough for a local LLM in n8n?

For small models and a proof of concept, often yes, for example with the Starter Kit on the CPU. For several users and large models like in the test, you need a lot of fast GPU or unified memory.

Do I still need a data processing agreement with a local LLM?

Not for the models themselves, as long as no data leaves your company. You may need one for hosting, remote maintenance or service providers with access to the system. Clarify this with your data protection officer.

Will my existing n8n workflows with OpenAI keep working after switching to local models?

Usually yes, if you swap the credential and turn off the Responses API in the Chat Model node. Still, retest every prompt, because a different model answers differently.

Sources

  1. n8n Blog: Run your n8n AI workflows on NVIDIA DGX Spark, with no cloud API
  2. NVIDIA: DGX Spark product page
  3. The Register: Nvidia raises DGX Spark price and adds a 64 GB version
  4. n8n: Self-hosted AI Starter Kit on GitHub
This article was created with AI.
Got a project in mind?

Tell us what you're planning. You'll get an honest assessment within 24 hours.

Got it! We will get back to you asap!
Whoopsie! Something went totally wrong here... Maybe just send an email to hi@blocksandcolors.de? Sorry!
Or email me directly: hi@blocksandcolors.de