Apple M1 Ultra

Apple M1 Ultra (128GB): the maximum-memory build of the M1 generation, and the only M1-era Mac that can hold a 235B-class MoE model's weights at all. It shares the 64GB Ultra's 20-core CPU, 800 GB/s LPDDR5 bandwidth, and UltraFusion dual-die design (two M1 Max dies fused with a 2.5 TB/s silicon interposer) — Apple doubled the memory pool without changing anything else, as a build-to-order upgrade on the Mac Studio that launched at $3,999 in March 2022.

120 GB of real headroom (after this site's standard 8 GB unified-memory reservation) unlocks a tier no smaller M1-family configuration reaches: this site's calculator fits Qwen3 235B-A22B, a 235-billion-parameter mixture-of-experts model, at Q2_K quantization and roughly 21.7 tok/s — the 64GB Ultra can't hold its weights at any quantization. On models that already fit both configurations, though, more memory buys quality rather than speed: Llama 3.3 70B runs at Q4_K_M (~14.1 tok/s) on the 64GB Ultra but at the higher-precision Q8_0 (~8.3 tok/s) here, since this site always recommends the best quantization that fits — extra headroom lets it pick a heavier format, which reads more bytes per token at the identical 800 GB/s bandwidth and so runs slower, not faster.

Same MLX and llama.cpp Metal support as the 64GB Ultra. Mixture-of-experts models like Qwen3 235B-A22B benefit less from Apple Silicon's bandwidth than a dense model of the same total size — llama.cpp and MLX still keep every expert's weights resident in memory even though only a fraction activate per token, so the capacity story here is about fitting the weights at all, not about MoE inference being unusually fast on Metal.

Apple M1 Ultra (64GB): the base memory configuration of Apple's first Ultra-tier chip — two M1 Max dies fused with Apple's UltraFusion interposer for 2.5 TB/s of chip-to-chip bandwidth, launched March 2022 exclusively in the Mac Studio desktop starting at $3,999. It doubles the M1 Max's 400 GB/s bandwidth to 800 GB/s while sharing the M1 Max 64GB config's memory ceiling.

Same 56 GB of real headroom as the M1 Max 64GB config — capacity is identical, so the same 56 of the 84 models tracked on this site fit natively, including Llama 3.3 70B and Qwen2.5 72B at Q4_K_M. What changes is speed: this site's calculator projects roughly 14 tok/s for both 70B-class models here, almost exactly double the M1 Max 64GB's ~7 tok/s, because doubling bandwidth doubles decode throughput on a workload this bandwidth-bound, without changing what fits.

Full MLX and llama.cpp Metal support. MLX in particular was built with Apple Silicon's unified-memory bandwidth in mind and tends to extract a larger share of the chip's rated bandwidth than llama.cpp's more general Metal backend — worth trying both if decode speed matters more than llama.cpp's broader format support.