n8n DGX Spark lets you run your AI workflows locally with no cloud API at all. On a single box with 128GB of unified memory you keep the whole stack loaded: chat, coding, vision, embeddings, reranking, STT, OCR, and image generation. With 273 GB/s memory bandwidth and a mixture-of-experts core, you get steady speeds and strict data control — every byte stays inside your building.
What are n8n AI workflows on DGX Spark?
They are ready-to-run n8n workflows for drafting emails, summarizing calls, reading invoices, answering questions from your own docs, and making images — without sending a byte to a cloud API.
You host a local model stack behind a single OpenAI-compatible endpoint that n8n’s OpenAI nodes already speak. Text-shaped jobs — summarizing, classifying, extracting, drafting, coding, and reading screenshots — go to one strong MoE model, while specialist models get a slot only when an LLM cannot do the job.
This solves privacy and sovereignty for clinics, law firms, and accountants: data stays on hardware you own, under your control, and you can show an auditor the exact path it took.
How to run n8n on DGX Spark?
In short: cache models offline, pin container versions, bring up the main model in vLLM, add the rest one by one, then put LiteLLM in front as a single OpenAI-compatible endpoint. Connect n8n via an OpenAI credential with a custom Base URL.
Step 1. Offline cache: download all models once and record their exact commits in a lock file. Then start each with --revision so you serve the weights you tested, not whatever main points to today.
Step 2. vLLM sanity check: don’t trust tags blindly. Inspect the version inside the container and use vllm/vllm-omni:v0.30.0 as the base for vLLM serve.
Step 3. Serve the backbone: the main Qwen3.6-35B-A3B model in NVFP4 starts under vLLM and becomes ready in about 311 seconds after docker run. Key flags include --gpu-memory-utilization 0.32, --kv-cache-dtype fp8, --moe-backend marlin, and --speculative-config with MTP.
Step 4. Fit the rest: the GPU and OS share a single 128GB pool (Linux may report around 121GB). Start services sequentially, wait for health checks, and drop the page cache before retrying a failed start, if needed.
Step 5. One endpoint: LiteLLM provides a single endpoint with named routes and virtual keys per team. In n8n, create an OpenAI credential, paste the LiteLLM virtual key, and set the Base URL to your LiteLLM /v1.
How does n8n DGX Spark perform under load?
Short answer: the 273 GB/s memory bandwidth plus MoE make text jobs fast for both single and concurrent users.
MoE activates only a few experts per token. Qwen3.6-35B-A3B has 35B parameters in total but about 3B active per step. In NVFP4, the active weights occupy far less than a dense model, so less is read from memory per token — more speed.
On DGX Spark, this model delivered about 120 tokens/s for one user and 359 tokens/s in total across eight users. Both figures include the reasoning tokens the model writes before answering.
vLLM selects Marlin kernels for NVFP4 GEMM. While math runs in 16-bit, you still get most of the gain: generation is memory-bandwidth bound, and 4-bit weights are a quarter of the size.
What models are in the stack?
One DGX Spark keeps the full service set in memory so workflows never touch the cloud. Every model allows commercial use:
- Chat/code/vision/extraction: nvidia/Qwen3.6-35B-A3B-NVFP4 Apache 2.0
- Embeddings: Qwen/Qwen3-Embedding-4B Apache 2.0
- Reranking: Qwen/Qwen3-Reranker-0.6B Apache 2.0
- Speech-to-Text: openai/whisper-large-v3-turbo MIT
- Document OCR: zai-org/GLM-OCR MIT
- Image generation: Tongyi-MAI/Z-Image-Turbo Apache 2.0
- Text-to-Speech: hexgrad/Kokoro-82M Apache 2.0
Cold-start load times and memory (after dropping the page cache): qwen3.6 ≈ 237 s and 43 GB, rerank ≈ 106 s and 8 GB, whisper ≈ 136 s and 8 GB, ocr ≈ 137 s and 11 GB, image ≈ 96 s and 20 GB, tts ≈ 21 s and 3 GB. With all seven models loaded, the box used about 108GB and had roughly 13GB free.
Remember: on shared memory, --gpu-memory-utilization is a fraction of the whole pool, and vLLM sizes its cache from what’s free at startup. Start services one at a time, watch the health checks, and retry if another process frees memory mid profiling.
n8n DGX Spark vs cloud API comparison
The DGX Spark stack processes everything on your own network. That simplifies contractual and regulatory posture: no data goes to a cloud API, and the path is audit-friendly.
Technically, you expose one OpenAI-compatible endpoint via LiteLLM with named routes per job. Models are pinned by commit, so upstream changes cannot swap weights or a license underneath you.
Performance stays predictable: one MoE model for all text, about 120 tokens/s per single user and 359 tokens/s total across eight users. Time to ready after start is around 311 seconds.
| Metric | Value | Note |
|---|---|---|
| Unified memory | 128GB | CPU+GPU shared (Linux sees ~121GB) |
| Memory bandwidth | 273 GB/s | GB10 Grace Blackwell |
| Main model size | 35B | ~3B active in MoE |
| Speed (1 user) | 120 tokens/s | includes reasoning tokens |
| Speed (8 users) | 359 tokens/s | aggregate |
| Time to ready | 311 s | after docker run |
Часті питання
How do n8n AI workflows run on NVIDIA DGX Spark?
You host a local stack on one box and expose a single OpenAI-compatible endpoint via LiteLLM. n8n talks to it like to OpenAI, but everything stays local: chat, coding, OCR, STT, TTS, embeddings, image generation. One MoE model in NVFP4 handles text-shaped jobs; specialist models run only when an LLM cannot cover the task.
n8n dgx spark vs cloud api comparison?
DGX Spark keeps data on your network with no cloud transfers. You pin weights by commit and can audit the data path. Speeds are around 120 tokens/s for a single user and 359 tokens/s across eight; time to ready after start is about 311 seconds.
How much is n8n dgx spark pricing?
The post doesn’t provide pricing. The stack runs locally, so there’s no per-token billing; LiteLLM tracks usage per team even if nothing is billed. The article does not include hardware or licensing costs.
Is n8n dgx spark for law firms a fit?
Yes, the scenario is framed for organizations with strict privacy and data sovereignty needs: clinics, law firms, and accountants. Models run on hardware you control with no cloud API. Use cases include transcribing calls, reading invoices, answering questions from your documents, and drafting replies.
How to connect n8n to DGX Spark?
Create a LiteLLM virtual key for the team, limited to the routes they need (e.g., qwen3.6, ocr). In n8n, add an OpenAI credential, paste the key, and set the Base URL to your LiteLLM /v1. For an Agent or a Basic LLM Chain, use an OpenAI Chat Model node with model By ID set to the route name and disable the Responses API.
Based on n8n Official Blog.
