Mac Studio M1 Ultra eyes bigger models

This is a capacity upgrade disguised as a shopping question: the chip matters, but the real unlock is having enough unified memory to keep larger models and longer contexts alive all day.

• –Apple’s M1 Ultra Mac Studio tops out at 128GB unified memory and 800GB/s bandwidth, which is why it keeps showing up in local-LLM conversations.
• –The replies naturally point toward 70B-ish instruction models, MoE checkpoints, and stacks like GGUF/llama.cpp or MLX, which is the right instinct once you stop optimizing for small-model demos.
• –For classification, summarization, and OSINT, the win is better conversational quality, more context, and a private always-on server, not just raw token speed.
• –The post captures the LocalLLaMA ethos well: spend on memory and silence, then build the stack around it.