Local AI
How to run open-weight AI models on your own hardware: what each model needs, and the machines that reach it.
-
12 Sep 2026
DeepSeek V4.1-Flash hardware requirements
510GB of weights, 8B active. A 2-bit MLX build runs on a Mac Studio M3 Ultra 256GB at 9.5 tokens per second — but not through llama.cpp, Ollama or LM Studio.
-
12 Sep 2026
GLM-5.3 hardware requirements
744B weights, 40B active per token, and a 4-bit build that wants 475GB. What the strongest open model needs, what you give up at each quantisation, and which machines reach it.
-
12 Sep 2026
How much memory to run an open model locally
Total parameters set the memory bill. Active parameters set the speed. A capacity ladder for the open models of 2026, from 17GB to 1.6TB.
