Apple M4 Max

Apple M4 Max (128GB): the maximum-memory M4 Max configuration, launched October 2024 in the MacBook Pro and, from March 2025, the Mac Studio, on the full, un-binned 16-core CPU (12P+4E)/40-core GPU die at 546 GB/s, the fastest memory bandwidth Apple has ever shipped in a laptop, and a real generational jump from the M3 Max's 400 GB/s ceiling.

With 120 GB of real headroom, this is the first M4 Max configuration to fit genuine frontier-scale MoE models: Qwen3 235B-A22B fits at Q2_K (102.05 GB, 14.8 tok/s) and GPT-OSS 120B reaches Q6_K (107.93 GB, 30.6 tok/s), both about 37% faster than the identically-sized M3 Max 128GB manages on the same fits (10.8 and 22.4 tok/s there), tracking the 546-vs-400 GB/s bandwidth gap almost exactly. Dense models do well too: Llama 3.3 70B and Qwen2.5 72B both reach Q8_0 (86.35 GB and 88.73 GB), near-full precision, at 5.7 and 5.5 tok/s. 73 of the 99 tracked models fit natively.

MLX and llama.cpp's Metal backend are both fully optimized here, and Apple Silicon's efficiency advantage over a discrete-GPU workstation is largest at this end of the lineup; this much memory bandwidth in a laptop chassis has no real equivalent on the discrete-GPU side.

Apple M4 Max (64GB): the mid-tier M4 Max configuration, available in the MacBook Pro and Mac Studio on the same full 16-core CPU/40-core GPU die as the 48GB and 128GB builds, all sharing the generation's 546 GB/s ceiling; only the 36GB base configuration ships on a slower, cut-down die (see this page's bandwidth note).

With 56 GB of real headroom, this is the first M4 Max tier where 70B-class dense models clear a real quantization instead of the aggressive Q2_K floor: Llama 3.3 70B reaches Q4_K_M (50.75 GB, 9.6 tok/s) and Qwen2.5 72B reaches Q4_K_M (52.12 GB, 9.4 tok/s). Command-R 35B goes further, hitting near-full-precision Q8_0 (53.70 GB, 9.1 tok/s). MoE models are faster still at this size: Qwen 3.5 35B-A3B fits at Q8_0 (41.86 GB, 40.5 tok/s). 63 of the 99 tracked models fit natively.

MLX and llama.cpp's Metal backend are both mature here. Nothing this configuration fits comes close to the default ~48 GB Metal working-set limit (75% of 64 GB), the largest fit here is under 54 GB, so the manual wired_limit override other Apple Silicon pages need doesn't bind at this capacity.

Apple M4 Max (48GB): the entry point into M4 Max's full, un-binned 16-core CPU/40-core GPU die, available in the MacBook Pro and Mac Studio, the same 546 GB/s as the 64GB and 128GB builds, a real step up from the cut-down 36GB base configuration's 410 GB/s.

With 40 GB of real headroom (identical to the M4 Pro 48GB configuration), this tier fits the same 58 of 99 tracked models, but at exactly double the speed on every shared fit: Llama 3.3 70B reaches the same Q2_K (32.88 GB) at 14.9 tok/s here versus 7.4 tok/s on the M4 Pro 48GB, and Qwen2.5 72B reaches 14.5 tok/s versus 7.3 tok/s, both track the 546-vs-273 GB/s bandwidth ratio almost exactly. MoE models do particularly well: Qwen 3.5 35B-A3B fits at near-full Q6_K (32.37 GB, 52.1 tok/s).

MLX and llama.cpp's Metal backend are both fully supported. Choosing this over the identically-priced-per-GB M4 Pro 48GB buys roughly 2x decode speed on every model both fit; the Max tier's advantage here is bandwidth, not capacity.

Apple M4 Max (36GB): the standard base configuration of Apple's M4 Max chip, launched October 2024 in the MacBook Pro and, from March 2025, the base Mac Studio, on a cut-down 14-core CPU (10P+4E)/32-core GPU die at 410 GB/s, 136 GB/s less than the 546 GB/s the 48GB, 64GB, and 128GB configurations get on the full die. Apple offers no way to pair 36GB with the faster die, the same binned-base-tier pattern the M3 Max shipped a generation earlier (see this page's bandwidth note).

With 28 GB of real headroom, this configuration fits the same 54 of 99 tracked models as the identically-sized M3 Max 36GB and M3 Pro 36GB (capacity decides what fits, not bandwidth), but runs about 37% faster than the M3 Max 36GB on every shared fit, since 410 GB/s is 37% more than that chip's 300 GB/s. Qwen3 32B reaches Q5_K_M (27.66 GB) at 13.3 tok/s here versus 9.7 tok/s on the M3 Max 36GB, and Qwen 3.5 35B-A3B reaches Q4_K_M (24.06 GB) at 52.4 tok/s versus 38.4 tok/s, both track the 410-vs-300 GB/s ratio closely. Even on this cut-down die, 410 GB/s outright beats the previous generation's full, non-binned M3 Max flagship die (400 GB/s).

MLX and llama.cpp's Metal backend are both mature here. The default macOS Metal working-set limit (~75% of 36 GB, 27 GB) sits just under this page's largest fit (Qwen3 32B's 27.66 GB Q5_K_M build), so raising it with `sudo sysctl iogpu.wired_limit_mb=28672` (leaving 8 GB for macOS) covers it with room to spare, the same fix and figure this site notes for the identically-sized M3 Max and M3 Pro 36GB configurations.