OpenClaw local stack fits M1 Max

64GB is enough to experiment, but not enough to make OpenClaw feel effortless; the real bottleneck is model quality and tool-use reliability, not just memory.

• –OpenClaw’s own guidance favors LM Studio or Ollama with a large model, and explicitly warns that small or aggressively quantized checkpoints degrade prompt-injection defenses.
• –A 32B Q4-class model is the sensible starting point; 70B Q4 may fit, but it will be slower and more cumbersome in an agent loop.
• –Fully local setups trade convenience for privacy: expect more latency, more retries, and more babysitting than you’d get from cloud models.
• –If the goal is dependable agent behavior rather than strict locality, a hosted fallback is the practical answer, but that conflicts with a local-only requirement.
• –For tinkering, light automation, and proof-of-concept workflows, Apple silicon with 64GB unified memory is a workable test bed.