For local LLMs with a Mac, rule of thumb is you always want an Ultra (due to memory bandwidth). Even an M1 Ultra is superior to an M6 Pro in this regard.

There are no configurations even close to running something comparable to frontier model variants, they're simply far too large, but something like full precision Qwen 35b or DeepSeek 70b at 50+ t/s is well within available configuration, and potential for plenty of room for large context sizes.

256GB is enough for DSv4 Flash, expect maybe ~30tg/s, and a lot better profile.

I'm using Flash heavily, and I would describe it as nearly as intelligent as Sonnet-class in agentic coding, but more usable. Less world knowledge of course, and definitely a bit less intelligent; but not _that_ much.

On usability: Takes less handholding, less likely to make unsolicited refactors or whatever, and the writing style is readable.

It's not great at super-long-horizon goals as the Claude 5 models are; but if you have a good harness, you can get around that.