OpenClaw hits performance wall on MacBook M4

The "painfully slow" local LLM experience on M4 is a classic case of context bloat meeting unoptimized inference backends.

• –Autonomous agents like OpenClaw inject 15K-20K tokens of workspace context into every request, crushing local models that lack prompt caching.
• –Falling back to CPU on M4 is a total performance killer; without proper Metal Performance Shaders (MPS) utilization, the chip's unified memory advantage is wasted.
• –Users should migrate to MLX-native frameworks or updated Ollama builds that leverage the M4's specialized Neural Engine and GPU cores.
• –For "free and fast" setups, small quantized models (e.g., Qwen2.5-7B or Llama-3.1-8B) at Q4_K_M quantization are the only viable path on base M4 hardware.
• –Real-world usability requires aggressive context management, including specific file excludes and frequent session resets to clear the token backlog.