AFAIK NVIDIA unified memory is not as... "unified" as Apple's

Yes, but

- the apps almost always allow you to remotely increase your stay - the apps almost always allow you to pay by the exact minute instead of by the quarter/half an hour

for the on-device model, yes it runs on the Neural Engine (at the moment) so a newer chip means faster, cheaper local inference. For the server side path this Claude package is about your machine is irrelevant since it's a network call. The same API covers both, so "best machine for AI" only bites when the session is actually local.

But we can imagine that the balance of what's on-device vs what's remote will move continuously towards the former as time, improved HW and improved local models keep progressing

it's cloud, the doc is explicit that requests go straight to api.anthropic.com with Apple not in the way.

so Claude via FM dies offline while Apple's on-device SystemLanguageModel (the ~3B one) keeps working. It isn't a hybrid really: the framework just has both implement the same LanguageModelSession protocol so "local 3B" and "remote frontier model" become a one-argument swap.

IMHO what's worth internalising is that the two share an API but nothing else: the on-device path runs on Apple's Neural Engine and costs battery (you can watch ANE power ramp while it works) while the cloud path costs API credits/tokens and does zero local compute. Same code, opposite cost model.

also: Fable takes 2× the usage of Opus

something I haven't seen highlighted anywhere yet, while I find it very interesting, is the distributed inference across Macs (JACCL over Thunderbolt 5), an OpenAI-compatible mlx_lm.server, agentic-on-Mac.

Apple keeps MLX (bring-your-own-weights) separate from Foundation Models / Core AI.

looks to me like the docs don't give a feature-parity table, but they do draw the "role" lines once you read across them:

- Core ML narrows to classic, non-neural ML (its own docs now point you there for "decision trees or tabular feature engineering")

- Core AI takes neural nets and transformers (the new .aimodel format, the new profiler)

- MLX stays the separate bring-your-own-weights track (its WWDC sessions draw no line back to Core AI at all)

coreai-opt is the successor to coremltools on the optimization side.

among the many things I found that were not highlighted anywhere (that I could find) is the distributed inference across Macs (JACCL over Thunderbolt 5), an OpenAI-compatible mlx_lm.server, agentic-on-Mac.

Apple keeps MLX (bring-your-own-weights) separate from Foundation Models / Core AI.

lol, it took me 48 hours to do (and re-do, and re-do) this test + write it up and now that I convinced myself to stop changing bits and just publish it... Google's just announced the Gemma 4 QAT models :-D

It would not change the core of my article since the bottleneck remains the memory bandwidth on the old M1 16GB though

you are making the mistake of confusing your experience, which is of course legit but anecdata, with "how it works" in general.

I'm an almost 50 years old Italian so not a spring chicken but I definitely learnt Italian growing up, not a dialect, and not "from school".

I guess it's the difference between growing up in a city vs a village.

Well yeah, GP's comment obviously only applies in the case that your native language is not standard Italian.

not obvious at all when every sentence uses "you" to indicate a general rule that applies to every Italian rather than "I" to indicate a personal experience