After a year of GPT-4 and Claude pilots in manufacturing, the honest question comes back: do we really have to ship process data to the cloud to get an MES assistant? In 2025–2026 the answer is no. Phi-4 (14B, Microsoft, MIT) at 4-bit quantization fits in 8 GB of VRAM, sqlite-vec gives you vector search in a single file with no server, and a Jetson Orin NX/AGX delivers 100–275 TOPS on the shop floor. This article walks through the concrete architecture, token-per-second benchmarks, 3-year TCO vs the OpenAI API, and what this means for AI Act, NIS2 and plant-level IT operations.

The Ollama project is an open, free software that allows running local AI models, specifically large language models (LLMs), on your own computer or server. Due to its low hardware requirements, this software is accessible to medium and small businesses, providing access to 93 AI models without significant limitations.