Comment by dofm

I am not — I'm using a 64GB M1 Max. Mostly with Muse Glimmer, sometimes Qwen 3.8 27B, Gemma 4 12B QAT, currently playing with LFM 2.5 24B-2B a bit.

I'm just saying, the base M6 is not very competitive in terms of memory bandwidth (the M5 Pro has about 80% faster memory bandwidth).

Compared with my old M1 Max, the M5 Pro is radically better at prefill, only about 15% slower at most at generation (and probably performs better at DFlash and MLX than the M1, so it might close the gap or even outperform the M1 Max)

So the models I am using/testing should all be quite fun, assuming the 48GB RAM machine. The Gemma 4 12B and LFM should run well on the 24GB machine. Neither of those are great coding models but they should be pretty good assistants. There are harsher quantisations of those other models that may fit.

Whether playing with these things locally is worth the extra spend, I don't know. It's probably more cost-effective to buy a cheaper Mac, experiment with the smaller models, and use Openrouter for the bigger ones.