Do any frameworks manage to use the neural engine cores for that?

Most stuff ends up running Metal -> GPU I thought

The thing about context/KV cache is that you can swap it out efficiently, which you can't with the activations because they're rewritten for every token. It will slow down as context grows (decode is often compute-limited when context is large) but it will run.

I strongly agree. People see local "GPT-4 level" responses, and get excited, which I totally get. But how quickly is the fall-off as the context size grows? Because if it cannot hold and reference a single source-code file in its context, the efficiency will absolutely crater.

That's actually the biggest growth area in LLMs, it is no longer about smart, it is about context windows (usable ones, note spec-sheet hypotheticals). Smart enough is mostly solved, combating larger problems is slowly improving with every major release (but there is no ceiling).

That should be covered by the harness rather than the LLM itself, no? Compaction and summarization should be able to allow the LLM to still run smoothly even on large contexts.

So prompt goes in 4x as fast but generates tokens slower.

I'd take that tradeoff. On my M3 Ultra, the inference is surprisingly fast, but the prompt processing speed makes it painful except as a fallback or experimentation, especially with agentic coding tools.

I run gpt-oss 120b model on ollama (the model is about 65 GB on disk) with 128k context size (the model is super optimized and only uses 4.8 GB of additional RAM for KV cache at this context size) on M4 Max 128 GB RAM Mac Studio and I get 65 tokens/s.

For chat type interactions prefill is cached, prompt is processed at 400tk/s and generation is 100-107tk/s, it's quite snappy. Sure, for 130,000 tokens, processing documents it drops to, I think 60tk/s, but don't quote me on that. The larger point is that local LLMs are becoming useful, and they are getting smarter too.

Please read the guidelines and consider moderating your tone. Hostility towards other commenters is strongly discouraged.

I'm not sure if you're just unaware or purposefully dense. It's absolutely possible to get those numbers for certain models in a m4 max and it's averaged over many tokens, I was just getting 127tok/s for 700 token response on a 24b MoE model yesterday. I tend to use Qwen 3 Coder Next the most which is closer to 65 or 70 tok/s, but absolutely usable for dev work.

I think the truth is somewhere in the middle, many people don't realize just how performant (especially with MLX) some of these models have become on Mac hardware, and just how powerful the shared memory architecture they've built is, but also there is a lot of hype and misinformation on performance when compared to dedicated GPU's. It's a tradeoff between available memory and performance, but often it makes sense.