MacBook Pro M3 Max tests local model stack

The right answer is almost certainly a tiered stack, not a single giant model: use a fast small model for routing and heartbeat jobs, a mid-size model for everyday agent work, and a larger reasoning model only when quality matters.

• –Apple’s M3 Max MacBook Pro tops out at 128GB unified memory and 400GB/s memory bandwidth, so 32B-class models are comfortable and 70B-class models are plausible in quantized form.
• –Ollama’s current library shows practical local picks like Qwen2.5 32B, DeepSeek-R1 32B, and Llama 3.3 70B, which maps well to a “small worker + larger thinker” setup.
• –For heartbeat, research, and automation, tool use matters more than raw model size; the model should orchestrate search, filesystem, and APIs rather than try to answer everything from weights alone.
• –A headless Mac is a good fit for a local runtime plus queue-driven jobs, because you can swap models without changing the automation layer.