What hardware do you need to run n8n AI agents on a local LLM?

Short answer: what hardware runs n8n with a local LLM?

For one person, an n8n AI Agent with tool calling runs on a 16 GB graphics card. For a small team, a 24 GB card such as a used RTX 3090 is the sweet spot. n8n itself is light: the model is what needs the memory. On a 16 GB card, gpt-oss-20b needs about 13.1 GB at standard quality (our estimate). On a 24 GB card (about 23.4 GB usable after our 0.6 GB reserve), Qwen3-Coder 30B A3B (20.2 GB) fits, and so does Qwen3.6-35B-A3B, the model n8n uses in its own DGX Spark guide (23.1 GB at UD-Q4_K_M, 8,192 tokens). A used RTX 3090 costs €1,246 / $1,395 (used median, pcprice.watch, 24 Oct 2026). n8n's DGX Spark reference stack keeps seven models loaded at once on 128 GB. You need that only if one box must serve chat, OCR, speech, embeddings and images to a whole team. All memory figures are localLLMrig estimates at 8,192 tokens, not benchmarks.

Before you buy a graphics card, test your agent prompt for free. Paste your n8n agent's system prompt and one real request into the free model board on localllmrig.com. Ask several open models the same question side by side, in the browser, with no login and no download, and see which one returns clean JSON for your tool call. Free models on 28 Oct 2026 include Gemma 4 26B A4B (18.1 GB, fits 24 GB) and NVIDIA Nemotron 3.5 Lightning (19.8 GB, fits 24 GB). Qwen3.6-35B-A3B and gpt-oss-20b are on the same board with credits. Test models free on localllmrig.com.

Why the model, not n8n, decides your hardware

n8n runs in Docker and calls a model server. The model server needs the memory: Ollama, in n8n's self-hosted AI starter kit, or vLLM or any OpenAI-compatible server, in n8n's DGX Spark guide. Memory is the model file plus conversation memory plus working space. Agent workflows send long prompts (the system prompt, the tool schemas and the memory), so plan for more context than plain chat.

n8n Self-Hosted AI Starter Kit Component Split

n8n + PostgreSQL + Qdrant = 2 CPU cores / 4 GB RAM / 20–50 GB SSD / 0 GB GPU. Ollama / vLLM takes the GPU VRAM, plus a concurrent webhook KV multiplier: each extra caller adds its own conversation memory. This split is our planning estimate, not a measured n8n benchmark.

Hardware tiers for n8n AI agents (8 GB to 128 GB)

Usable memory is from our hardware pages. On graphics cards we keep 0.6 GB free. On Macs and the Ryzen AI Max+ 395 mini PC it is the figure on each machine page. The role column is a guide to what the memory can hold. It is not a speed test.

Model pages for the tiers above: Qwen2.5 14B, LFM3-1.7B, Spark-X2.6-4B, Llama 3.1 8B, gpt-oss-20b, Mistral Small 3.2 24B, Qwen3 Coder 30B A3B, Qwen3.8-27B, Gemma 4 26B A4B, Qwen3.6-35B-A3B, gpt-oss-120b. Memory tiers: 8 GB, 16 GB, 24 GB, 32 GB, 128 GB. Machines: GeForce RTX 3060 12GB, Radeon RX 9060 XT 16GB, GeForce RTX 4060 Ti 16GB, Mac mini M4 16GB, GeForce RTX 3090 24GB, Used PC with RTX 3090 24GB, GeForce RTX 5090 32GB, GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB, Mac Studio M5 Max 128GB, NVIDIA DGX Spark 128GB.

Before you buy a graphics card, test your agent prompt for free. Paste your n8n agent's system prompt and one real request into the free model board on localllmrig.com. Ask several open models the same question side by side, in the browser, with no login and no download, and see which one returns clean JSON for your tool call. Free models on 28 Oct 2026 include Gemma 4 26B A4B (18.1 GB, fits 24 GB) and NVIDIA Nemotron 3.5 Lightning (19.8 GB, fits 24 GB). Qwen3.6-35B-A3B and gpt-oss-20b are on the same board with credits. Test models free on localllmrig.com.

Tool calling: what the model must support

The AI Agent node decides which tools to call, and it needs at least one tool connected (n8n docs). Since n8n 1.82.0 every AI Agent works as a Tools Agent, which uses LangChain's tool-calling interface. Pick a model that supports tool calling in your runtime. On ollama.com, the library pages for llama3.1, qwen3, qwen3.6, qwen3-coder, gpt-oss, mistral-small3.2, gemma4 and lfm3 carry the "tools" tag (checked 28 Oct 2026). On vLLM, n8n's guide warns that the reasoning parser and the tool-call parser must match the model family, and to check those flags before you blame the model. For fixed-format extraction, turn thinking off: in n8n's test, extraction took 0.59 seconds with thinking off and 12.1 seconds with thinking on.

The n8n docs list OpenAI, Groq, Mistral Cloud, Anthropic and Microsoft Foundry as supported chat models for the Tools Agent. That list does not name Ollama. n8n's own 22 Oct 2026 post shows the Ollama Chat Model connected to the AI Agent. This page does not call Ollama an officially supported Tools Agent model. It says what n8n shows.

Qwen3.6-35B-A3B on 24 GB vs n8n's DGX Spark (128 GB)

A single user or a small team running one text model for agent workflows gets Qwen3.6-35B-A3B on a $1,395 used 24 GB card (about 20% of a DGX Spark's $6,950). A complete used RTX 3090 PC (€2,346 / $2,413, our build estimate, 24 Oct 2026) is about 35% of that $6,950. The 128 GB box earns its price when one machine must keep OCR, speech, embeddings, reranking and image models loaded next to the LLM. Read the used RTX 3090 buying guide, RTX Spark, Mac Studio and Strix Halo, Sunghoon Baek on one DGX Spark, and OscarJHernandez on a GMKtec EVO-X2 128 GB.

Context length: why 24 GB gets tight for agents

Conversation memory for Qwen3.6-35B-A3B comes from its layout: 10 full-attention layers, 2 key-value heads, head size 256, FP16. That is 0.17 GB at 8,192 tokens on the model page. File sizes are the Unsloth sizes checked 24 Oct 2026, plus 0.80 GB of working space. All figures in this table are estimates, not measured runs.

n8n's own vLLM setup uses a maximum model length of 65,536 tokens. For agent workflows on a 24 GB card, use UD-IQ4_XS (a smaller download, answers a little less well) or a 32 GB card. See what the quality names mean.

Mac and mini PC setups

n8n's starter kit says a Mac can't expose your GPU to the Docker instance. Run Ollama natively on the Mac and point n8n at http://host.docker.internal:11434 (n8n's Ollama docs). Usable memory: 10.7 GB on a Mac mini M4 16GB, 21.3 GB on a Mac mini M6 32GB, and about 96 GB on 128 GB Macs such as the Mac Studio M5 Max 128GB and on Ryzen AI Max+ 395 mini PCs such as the GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB (96 GB). We do not rate which models fit the NVIDIA DGX Spark 128GB.

What the n8n 3.0 beta changes for local AI workflows

n8n 3.0.0 is a beta, released 27 Oct 2026. n8n says the stable release comes in early December. 3.0 adds no new local-model feature. It is a cleanup and safety release. Local model support comes from nodes that already exist in 2.x: Ollama Chat Model, OpenAI Chat Model with a custom base URL, and the AI Agent (Tools Agent).

• 3.0 focuses on reliability, stability and security. New features ship on 2.x and 3.x as they are ready. n8n counts more than 45 releases since 2.0 (December 2025). Source: Introducing n8n 3.0.
• Self-hosted n8n 3.0 is Docker-only. npm and npx n8n installs are no longer supported. There is a one-line setup script. Kubernetes and Helm are unaffected. Source: n8n 3.0 breaking changes.
• AI Agent node v1 is removed, together with the SQL, Conversational, OpenAI Functions, Plan and Execute, and ReAct agent modes. The current version uses Tools Agent. Also removed: the legacy OpenAI node, OpenAI Assistant, the OpenAI Model node (replace it with OpenAI Chat Model), the legacy HTTP Request Tool, LangChain Code (legacy), and the Manual Chat Trigger. Chat Hub is off by default.
• With SSRF protection on, the default block list covers localhost and private networks. 3.0 adds four more ranges, including 100.64.0.0/10 (CGNAT, used by Tailscale and similar networks). If n8n calls a model server on such an address, allow-list it. The settings are N8N_SSRF_ALLOWED_IP_RANGES and N8N_SSRF_ALLOWED_HOSTNAMES.
• 2.x gets security and critical fixes for 12 months after stable 3.0. 1.x support ends 31 Dec 2026. Upgrading to 3.x is one-way, with no rollback. Run the Migration Report in n8n 2.39 or newer before you upgrade.

Connect n8n to your local model (3 options)

• Ollama Chat Model node, plus an Ollama credential with a base URL. On Docker, use host.docker.internal, or the container name if both run in Compose. The Ollama Chat Model docs still mention only Llama 2 in the model dropdown. That page is out of date. Do not treat it as the model list.
• OpenAI Chat Model node with a custom base URL to vLLM, LiteLLM or a llama.cpp server. Set the model by ID and turn off "Use Responses API". n8n's DGX Spark guide says the node uses the Responses API by default from node version 1.3. n8n also says LiteLLM must be 1.84.0 or newer, because two PyPI releases were backdoored in March 2026.
• HTTP Request node for fields the chat-model nodes do not set, such as images or chat template options.

Buy, rent, or test first?

Compare owning a machine with renting or paying per use on buy or rent. For a team, local LLM for business in the EU lists the questions that stay open after the model runs on your hardware. Open the calculator for this model at standard quality: Qwen3.6-35B-A3B on the calculator. Dollar amounts next to a euro price are our conversion of that euro figure, dated with the price.

Before you buy a graphics card, test your agent prompt for free. Paste your n8n agent's system prompt and one real request into the free model board on localllmrig.com. Ask several open models the same question side by side, in the browser, with no login and no download, and see which one returns clean JSON for your tool call. Free models on 28 Oct 2026 include Gemma 4 26B A4B (18.1 GB, fits 24 GB) and NVIDIA Nemotron 3.5 Lightning (19.8 GB, fits 24 GB). Qwen3.6-35B-A3B and gpt-oss-20b are on the same board with credits. Test models free on localllmrig.com.

Sources

• n8n blog, Introducing n8n 3.0 (27 Oct 2026)
• n8n docs, v3.0 breaking changes
• n8n blog, Run n8n AI workflows on NVIDIA DGX Spark (19 Oct 2026)
• n8n blog, Self-hosted LLM vs cloud APIs (22 Oct 2026)
• n8n docs, AI Agent
• n8n docs, Tools Agent
• n8n docs, Ollama Chat Model
• n8n docs, Ollama Docker networking
• n8n self-hosted AI starter kit
• Ollama library, qwen3.6 tools tag (checked 28 Oct 2026)
• pcprice.watch, RTX 3090 used price
• NVIDIA Marketplace, DGX Spark