Mac Mini M4 Pro and OpenClaw: Testing Local LLM Inference

The author tests local LLM inference on a Mac mini M4 Pro with 48 GB of RAM, using OpenClaw as an AI agent and LM Studio to run models. OpenClaw is a self-hosted gateway for connecting applications and messengers with the ability to integrate AI agents. Three models were selected for testing: google/gemma-4-26b-a4b-qat (MLX, 4-bit), qwen/qwen3.6-35b-a3b (GGUF, Q4_K_M), and google/gemma-4-31b (GGUF, Q4_K_M). Ping times were 4, 3, and 12 seconds, respectively. When asked for news about Selectel, the qwen3.6-35b-a3b model gave the most detailed response in 60 seconds, gemma-4-26b-a4b-qat delivered an analytical response in 60 seconds, and gemma-4-31b provided a brief response in 300 seconds. Synthetic Anubis benchmarks showed that gemma-4-26b-a4b-qat generates 59 tokens per second with an energy consumption of 0.38 Joules per token, qwen3.6-35b-a3b produces 50 tokens per second and 0.45 Joules per token, while gemma-4-31b manages only 10 tokens per second and 2.73 Joules per token. GPU utilization reached 99%, while CPU usage was below 10%. The MoE models (gemma-4-26b and qwen3.6-35b) are significantly faster and more energy-efficient than the dense gemma-4-31b model.