The price of an AI assistant for a small business is driven by setup work plus the monthly LLM traffic, and self-hosting a model to shrink that monthly bill rarely pays off below steady, heavy use. n8n just released a routing framework that explains when on-prem wins and when a cloud API stays the cheaper option.
The piece walks through open-weight models you can actually run locally - Phi-4-mini, Qwen3 8B, Gemma 3 12B, DeepSeek R1-Distill-Qwen-14B, Mistral Small 3.1, and Llama 4 Scout - with VRAM requirements and quantization notes. The core idea in the n8n blog post is hybrid routing: cheap and non-sensitive requests go to the cloud, regulated or high-volume stable traffic goes to a model you host.
What this means for a small business
For a typical SMB running leads, invoices, or support bots, a single GPU costs more than a year of Claude or GPT calls, so cloud wins by default. Self-hosting starts to make sense in two situations: regulated data that cannot leave the perimeter, and workloads with constant heavy volume where API bills become the dominant line item. In the legal document pipeline I built, local inference was chosen for the first reason - the files cannot sit on a foreign cloud. For a 20-person e-commerce team doing 500 chat replies a day, I would route everything through a cloud model and skip the GPU entirely.
Why it matters: Pick infrastructure after you know the request volume and data constraints, not before - most small operations end up paying for hardware they never fully use.
Source: n8n Blog