Apple M2 Max

Apple M2 Max (96GB): the flagship memory configuration of Apple's second-generation Max chip, pairing the 12-core CPU (8 performance + 4 efficiency) with the full 38-core GPU. It shares the same 400 GB/s of LPDDR5-6400 bandwidth as the 32GB and 64GB configs — Apple never cut bandwidth across the M2 Max's memory range the way it later did with the 96GB M3 Max's binned die.

88 GB of real headroom after this site's standard 8 GB reservation unlocks a different tier of model than the 64GB sibling: Llama 3.3 70B fits at Q8_0 (86.4 GB, ~4.2 tok/s) instead of Q4_K_M, and GPT-OSS 120B — a model that doesn't fit the 64GB configuration at any quantization, since even its smallest real build needs about 63 GB before KV cache — fits here at Q4_K_M (80.2 GB) around 30 tok/s. Qwen 3.5 122B-A10B, a mixture-of-experts model, reaches Q4_K_M (85.6 GB) at roughly 14.3 tok/s. 61 of the 84 tracked models fit natively, 5 more than the 64GB configuration.

MLX and llama.cpp's Metal backend are both mature on this chip; community benchmarks commonly report MLX running 20-87% faster than llama.cpp Metal for dense models under 14B parameters, narrowing toward parity above roughly 27B parameters where decode is fully bandwidth-bound. Several of the larger fits above — Llama 3.3 70B at Q8_0 (86.4 GB) and GPT-OSS 120B at Q4_K_M (80.2 GB) among them — exceed macOS's default ~72 GB Metal working-set limit (roughly 75% of unified memory), so raising it by hand with `sudo sysctl iogpu.wired_limit_mb=90112` (leaving 8 GB for macOS, matching this page's reservation) is necessary before loading them.

Apple M2 Max (64GB): The M2 Max pairs a 12-core CPU (8 performance + 4 efficiency) with up to a 38-core GPU. Unlike the M3 Max generation that replaced it — where the 96GB build bins down to a slower die — every M2 Max memory size (32GB, 64GB, 96GB) shares the same 400 GB/s of LPDDR5-6400 unified memory bandwidth regardless of GPU core count; Apple never cut bandwidth on this chip.

This calculator always recommends the highest-quality quantization that still fits within this configuration's roughly 56 GB of usable memory (64 GB minus 8 GB reserved for macOS). For 32B-class dense models — Qwen3 32B, Qwen 2.5 32B and its Coder variant, and DeepSeek-R1-Distill-Qwen-32B — that's Q8_0: Qwen3 32B lands at 40.55 GB and roughly 8.8 tok/s, with the other three landing within half a gigabyte and a tenth of a token per second of that (41.1 GB, 8.7 tok/s). Hand-picking Q4_K_M instead (23.88 GB for Qwen3 32B) nearly doubles that to about 15 tok/s at some quality cost. Dense 70B-class models fit at Q4_K_M: Llama 3.3 70B reaches 50.75 GB at roughly 7.1 tok/s, and Qwen 2.5 72B is the largest dense model this configuration fits natively at all, at 52.12 GB and 6.9 tok/s — both track a widely-cited real-world report of about 7 tok/s for Llama 2 70B Q4 on an M2 Max. MoE models are where the extra bandwidth pays off, since decode only reads the active parameters: Qwen3-30B-A3B and Qwen 3.5 35B-A3B both fit at full Q8_0 (36.62-41.86 GB) around 28-30 tok/s. GPT-OSS 120B is the exception that proves the rule — it doesn't fit this configuration at any quantization, because OpenAI never released its MoE expert weights above native MXFP4: even the smallest real build is about 63 GB before KV cache and overhead, past the ~56 GB of usable memory here regardless of which quant label you pick. Qwen 3.5 122B-A10B, the largest model this configuration fits at all, reaches Q2_K (54.47 GB) around 21.6 tok/s.

MLX and llama.cpp's Metal backend are both mature on this chip. Community benchmarks commonly report MLX running 20-87% faster than llama.cpp Metal for dense models under 14B parameters, with the gap narrowing toward parity above roughly 27B parameters, where decode is fully bandwidth-bound and the software stack matters less. One gotcha specific to this 64GB configuration: macOS's Metal driver only exposes about 75% of unified memory to the GPU by default — around 48 GB here — enough for every 32B-class Q8_0 model above but not the 50.75 GB Llama 3.3 70B Q4_K_M this page recommends. Raise it by hand with `sudo sysctl iogpu.wired_limit_mb=57344` (leaving 8 GB for macOS, matching what every estimate on this page already assumes) before loading anything close to that ceiling.

Apple M2 Max (32GB): the base memory configuration of Apple's second-generation Max chip, launched January 2023 alongside the M2 Pro on the 14-inch and 16-inch MacBook Pro. It shares the 12-core CPU (8P+4E) with every other M2 Max memory size and doubles the M2 Pro's 200 GB/s bandwidth to 400 GB/s from the same starting memory pool.

24 GB of real headroom after the standard 8 GB reservation — identical capacity to the M2 Pro 32GB configuration, so the same 46 of 84 tracked models fit natively, including Qwen3 32B at Q4_K_M (23.9 GB) and Mixtral 8x7B at Q2_K. What changes is speed: this site's calculator projects roughly 15 tok/s for Qwen3 32B here, about double the M2 Pro 32GB's 7.5 tok/s, and mid-size models get real precision headroom too — Llama 3.1 8B and Qwen2.5 7B fit at full BF16 (19.1 GB and 17.6 GB) at roughly 18.7-20.4 tok/s rather than being squeezed onto a lower quant.

Full MLX and llama.cpp Metal support, same as the 64GB and 96GB siblings. Because macOS's Metal working-set reservation stays roughly fixed regardless of installed memory, it eats a much smaller share of this 32 GB pool than it does on a 16 GB Apple Silicon configuration, leaving comfortable headroom for every fit above.