What LLMs can the Apple M3 Ultra (512GB) run?

62 of the 62 open models in this roster have an on-device configuration with usable context. The largest by exact parameter count is Hy3 at Q8_0, with room for 256k tokens. “Largest” describes parameter count, not model quality or task performance.

The device publishes 512 GB of LPDDR5 unified. This calculator budgets 75%, or 384.0 GiB, for model weights and KV cache; the remainder is an explicit allowance for the runtime, driver or operating system, workspace, and display. It is an assumption rather than a vendor specification.

No. They are bandwidth-bound roofline ceilings: 819 GB/s divided by the bytes one decoded token actually reads, which the table prints beside every ceiling. That is not the size of the file on disk — a token reads the weights it routes through, all of them for a dense model but only the selected experts for a mixture-of-experts, plus one pass over the KV cache. gpt-oss 120B holds 60.8 GiB resident at As released yet reads 3.3 GiB/token, which is what produces its 234 tok/s. Real throughput is lower because kernels, cache traffic, prompt processing, scheduling, and runtime overhead also consume time.