Skip to main content
Rinda Logo

You Can Train LLMs Without NVIDIA — The Quiet Revolution of Swift + Metal

Apple's M2 Ultra runs a 70B parameter model at ~20 tok/sec via MLX — no NVIDIA, no cloud bill. Swift+Metal is quietly making a $3,499 Mac Studio a serious LLM tool.

GRINDA AI
May 11, 2026
1 min read
Share

Running your AI SDR locally on M2 Ultra (192GB, ~$3,499) means zero API latency and zero data leakage. 70B models at ~20 tok/sec via MLX (int4 quantized). The real moat in AI sales isn't GPU access — it's proprietary CRM data no one can scrape. What's your local inference stack?

Local LLMApple MLXOn-Premise AIM2 UltraRinda