Apple M2 Ultra

Apple M2 Ultra (192GB): the maximum-memory configuration of Apple's second Ultra-tier chip — two M2 Max dies fused with Apple's UltraFusion interposer, a high-speed silicon bridge that lets the pair act as one chip, launched June 2023 in the Mac Studio and Mac Pro. 192GB is the real ceiling for this chip: Apple never released the 384GB "M2 Extreme" that pre-launch leaks predicted, so 192GB — not 384GB — is the most memory any Mac has ever shipped with an M2-generation chip.

184 GB of real headroom after this site's standard 8 GB reservation clears every dense model this site tracks up to 72B at full BF16 precision — Llama 3.3 70B and Qwen2.5 72B both fit at BF16 (159.8 GB and 164.3 GB) instead of being squeezed into Q4, at roughly 4.4-4.5 tok/s. It's also the only real M2 Ultra configuration that reaches 405B-parameter territory: Llama 3.1 405B fits at Q2_K (177.6 GB, ~4 tok/s), a capability the 64GB base configuration doesn't have at any quantization. Mixture-of-experts models are where the extra memory pays off most: Qwen3 235B-A22B fits at Q4_K_M (162.1 GB) around 13.8 tok/s, and GPT-OSS 120B reaches Q8_0 (139.6 GB) at roughly 34.8 tok/s. 71 of the 84 tracked models fit natively.

Full MLX and llama.cpp Metal support. Community llama.cpp benchmarks on 76-core-GPU M2 Ultra report roughly 88-94 tok/s decode on a 7B model at Q4_0, and 1,100-1,400 tok/s prompt processing at F16 (ggml-org/llama.cpp Discussion #4167) — figures set by GPU core count and 800 GB/s bandwidth, independent of which real memory configuration (64GB or 192GB) is installed.

Apple M2 Ultra (64GB): the base memory configuration of Apple's second Ultra-tier chip, the entry point into the Mac Studio and Mac Pro's Ultra lineup starting at $3,999 in June 2023. It shares the 192GB configuration's full 800 GB/s of unified memory bandwidth and 24-core CPU (16P+8E) — Apple doesn't cut bandwidth for the smaller memory build on this chip, unlike the bandwidth-binned 96GB M3 Max a generation later.

56 GB of real headroom after the standard 8 GB reservation — identical capacity to the M2 Max 64GB configuration, so the same 56 of 84 tracked models fit natively, including Llama 3.3 70B and Qwen2.5 72B at Q4_K_M. What changes is speed, not capacity: this site's calculator projects roughly 14.1 tok/s for Llama 3.3 70B here, almost exactly double the M2 Max 64GB's 7.1 tok/s, because doubling bandwidth (800 vs 400 GB/s) doubles decode throughput on a workload this bandwidth-bound without changing what fits.

Same MLX and llama.cpp Metal maturity as the 192GB sibling. The Ultra tier's advantage over a same-capacity Max configuration is decode speed, not new capability — choosing this over an M2 Max 64GB buys roughly 2x tokens/sec on models that already fit both, not the ability to run anything new.