Apple M3 Max

Apple M3 Max (128GB): the flagship M3 Max configuration, launched November 2023 on the full, un-binned die: 16-core CPU (12P+4E), up to a 40-core GPU, and the generation's fastest 400 GB/s bandwidth — the same ceiling the 48GB and 64GB builds share, well ahead of the binned 36GB/96GB builds' 300 GB/s (see this page's own bandwidth note for why those two memory sizes bin down). Even at full bandwidth, this is still a single M3 Max die — see the bandwidth note for how far short that leaves it of the dual-die M3 Ultra.

With 120 GB of real headroom, this is the first M3 Max configuration to fit genuine frontier-scale MoE models: Qwen3 235B-A22B fits at Q2_K (102.05 GB, 10.8 tok/s), and MiniMax M2.5/M2.7 (229B) fit at the same quant (100.05 GB, 21.7 tok/s) — both need an M3 Ultra 256GB configuration to run at a meaningfully better quant. On models both the 96GB and 128GB builds already fit, the extra bandwidth over the binned 96GB die shows up directly: Llama 3.3 70B holds at the same Q8_0 (86.35 GB) on both, but reaches 4.2 tok/s here versus 3.1 tok/s on the 96GB build. GPT-OSS 120B goes further, moving up from Q4_K_M on the 96GB build to Q6_K here (107.93 GB) at 22.4 tok/s — nearly the same speed as the 96GB build's lower-precision Q4_K_M fit (22.5 tok/s), because the extra bandwidth roughly pays for the bigger, higher-precision file.

MLX and llama.cpp's Metal backend are both mature here. Like the 96GB and 36GB binned builds, the macOS Metal default working-set limit (~75% of unified memory, 96 GB on this 128GB build) sits below the biggest fits this page reports — including the 102.05 GB Qwen3 235B-A22B fit above — so raising it with `sudo sysctl iogpu.wired_limit_mb=122880` (120 GB, leaving 8 GB for macOS) is worth doing before loading anything past the ~96 GB default.

Apple M3 Max (96GB): late-2023 laptop chip, and the 96GB build isn't the M3 Max flagship — Apple only pairs 96GB unified memory with the binned die (14-core CPU, 30-core GPU), which has 8 of 32 memory controllers disabled. That caps bandwidth at 300 GB/s, a real 25% cut from the 400 GB/s every 48GB, 64GB, and 128GB M3 Max ships with. It's the one Apple Silicon generation where memory bandwidth doesn't scale cleanly with capacity.

This calculator always recommends the highest-quality quantization that still fits, which for a dense 70B model here is Q8_0 (86 GB) — technically native, but only about 3 tokens/sec, too slow for real back-and-forth. Drop to Q4_K_M (51 GB) by hand and Llama 3.3 70B reaches roughly 5 t/s; the 10-15 t/s often quoted for "70B Q4 on M3 Max" comes from the faster 400 GB/s configurations, not this 300 GB/s one. Mid-size dense models do better: Qwen 3.6 27B fits at full BF16 (61 GB, ~4.4 t/s) or Q4_K_M (19 GB, ~14 t/s). MoE is where the extra memory pays off — Qwen3-30B-A3B and Qwen 3.5 35B-A3B both fit at full BF16 (~68-79 GB) around 11-12 t/s, and GPT-OSS 120B fits at Q4_K_M (80 GB) at roughly 22-23 t/s, since decode only reads its ~5B active parameters. Frontier MoE models over 350B don't fit at any quantization here — GLM 4.5 (355B) needs an M3 Ultra with at least 256GB, and DeepSeek V3 (671B) or Kimi K2.6 (1T) need the 512GB tier.

MLX and llama.cpp's Metal backend are both mature on this chip; community benchmarks commonly put MLX-LM 10-30% faster for sustained generation, though your mileage will vary by model and quant. One real gotcha at this memory size: macOS's Metal driver only exposes about 75% of unified memory to the GPU by default — around 72 GB on a 96GB machine — so anything near the ceiling (like the 86 GB Q8_0 Llama 3.3 70B above) won't load until you raise it with `sudo sysctl iogpu.wired_limit_mb=90112` (~88 GB, leaving 8 GB for macOS). Routine for anyone running large local models on Apple Silicon, but an easy thing to trip over the first time.

Apple M3 Max (64GB): the 64GB full-die M3 Max configuration, launched November 2023 with the same 16-core CPU (12P+4E) and up to 40-core GPU as the 48GB and 128GB builds, all sharing the un-binned die's 400 GB/s bandwidth ceiling — the binned die only shows up at 36GB and 96GB (see this page's bandwidth note).

With 56 GB of real headroom, this is the M3 Max configuration where 70B-class dense models properly clear Q4_K_M: Llama 3.3 70B reaches 50.75 GB at 7.1 tok/s and Qwen2.5 72B reaches 52.12 GB at 6.9 tok/s — identical figures to what this site's calculator projects for the equally-56 GB, equally-400-GB/s M1 Max 64GB and M2 Max 64GB configurations, since bandwidth at this specific capacity hasn't moved across three Apple Silicon generations. 56 of the 84 tracked models fit natively — the same count as those two earlier chips at the identical capacity.

MLX and llama.cpp's Metal backend are both mature here. This configuration's default macOS Metal working-set limit (~75% of 64 GB, 48 GB) sits below the 50.75 GB Llama 3.3 70B fit above — raise it with `sudo sysctl iogpu.wired_limit_mb=57344` (56 GB, leaving 8 GB for macOS) before loading it, the same fix and figure this site notes for the identically-sized M1 Max and M2 Max 64GB configurations.

Apple M3 Max (48GB): the entry M3 Max configuration, launched November 2023 on the same full 400 GB/s die as the 64GB and 128GB builds — Apple's cheapest way into M3 Max, one memory tier above where the binned die tops out at 96GB.

With 40 GB of real headroom, this is the smallest M3 Max configuration that touches 70B-class models at all — Llama 3.3 70B, Llama 3.1 70B, and the DeepSeek-R1 70B distill all fit at Q2_K (32.88 GB, 10.9 tok/s), an aggressive quant that trades real quality for capacity. MoE models fare much better at this size: Qwen3-30B-A3B fits at full Q8_0 (36.62 GB, 28 tok/s) and GPT-OSS 20B fits at Q8_0 (25.23 GB, 24.7 tok/s), both essentially full-precision, since decode only reads their few billion active parameters rather than the whole file. 51 of the 84 tracked models fit natively.

MLX and llama.cpp's Metal backend are both mature here. This capacity is exactly where the macOS Metal default working-set limit (~75% of 48 GB, 36 GB) starts to matter even for models that comfortably fit in the full 40 GB pool: Llama 3.1 8B's best-fitting quant here is FP32 (37.04 GB), just over that default ceiling, so raising it with `sudo sysctl iogpu.wired_limit_mb=40960` (leaving 8 GB for macOS) is worth doing even at this modest capacity, not just on the larger builds.

Apple M3 Max (36GB): the entry-level M3 Max configuration, and — like the 96GB build — pairs with the cut-down 14-core CPU (10P+4E)/30-core GPU die, which has 8 of 32 memory controllers disabled. That caps bandwidth at 300 GB/s rather than the 400 GB/s every 48GB, 64GB, and 128GB M3 Max ships with (see this page's bandwidth note).

With 28 GB of real headroom, this configuration fits the exact same 47 of the 84 models as the identically-sized 36GB M3 Pro — capacity, not bandwidth, decides what fits — but at roughly double the speed on every model both share, since 300 GB/s is double the M3 Pro's 150 GB/s. Qwen 3.5 35B-A3B reaches Q4_K_M (24.06 GB) at 38.4 tok/s here versus 19.2 tok/s on the M3 Pro 36GB, and Qwen3 32B reaches Q5_K_M (27.66 GB) at 9.7 tok/s versus 4.9 tok/s — both almost exactly 2x.

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.