upvote
deleted
reply
Don't worry, there are exactly zero Apple chips that are good for any sort of local inference right now. Unified ram is fast for ram, but dogshit slow compared to actual dedicated VRAM on graphics cards.
reply
Your statement is not really true any more.

I have an old M1 Max with 32GB that runs inference models fine, especially those using MLX.

Dedicated graphics cards are faster but not to the point that it matters.

That’s a six year old machine.

My newer machine M5 max 128GB will far outperform your typical 32GB gaming card once the model exceeds memory.

To even get close to that on a dedicated graphics card you are paying upwards of $20K.

That said, models are getting smaller and faster which allows me to run multiple different models with ease.

[edit] Just checked the 5 models I use take up 57GB when all loaded at the same time.

reply
The $8k machine will out preform the 1k gaming machine sure but it wont out preform anything in a similar price range.
reply
> Dedicated graphics cards are faster but not to the point that it matters.

It definitely matters, if you're intending to run a Claude Code/OpenCode style agent workflow. Most of those harnesses start with 8-12k token contexts, which is a lot of prefill for a Mac but cheap for a CUDA GPU. The GPGPU compute on the fastest Macs is still trailing behind Nvidia's laptop GPUs; the highest-bandwidth Apple Silicon chip (now the M5 Ultra) has ~7x lower memory bandwidth than a single B100 card.

There's a good reason why Apple Silicon isn't to be found anywhere in the datacenter buildout. It's nonviable for training, and wastes electricity running real-world inference workloads.

reply
Is it too much to give the harness a moment to load up the context when starting a new session? Is the brief pause truly a deal breaker?
reply
[dead]
reply
That selling point is probably more about using a claude code subscription than running your own models. Because of apple's lockin there are certain things you can only do on apple devices and a mac mini is useful for that
reply