Apple M3 Ultra

Apple M3 Ultra (512GB): the maximum-memory Mac Studio configuration Apple introduced in March 2025, and the only M3 Ultra memory tier that requires the 32-core CPU (24P+8E)/80-core GPU upgrade die — Apple never offered 512GB on the base 28-core/60-core chip. M3 Ultra is two complete M3 Max dies fused with UltraFusion, a 2.5 TB/s silicon interposer (184 billion transistors total) — the same packaging trick Apple has used for every Ultra chip since the M1 generation — which is why its 819 GB/s bandwidth tracks roughly double a single M3 Max die rather than a new memory architecture.

With 504 GB of real headroom after this site's standard 8 GB reservation, this configuration fits 82 of the 84 models tracked on this site natively in VRAM. DeepSeek V3 and DeepSeek R1 (671B parameters) both fit at Q4_K_M (458.25 GB, 8.7 tok/s), and Kimi K2.6 — a full 1 trillion parameters — fits at Q2_K (429.73 GB, 15.1 tok/s), one of the very few ways to hold a trillion-parameter model's weights on a single consumer-purchasable machine at all. GLM 4.5 and GLM 4.6 (355B) run at Q8_0 (426.11 GB, 5.6 tok/s) here, full-precision headroom the smaller 256GB tier can't offer them.

MLX and llama.cpp's Metal backend both run frontier MoE models here, but neither parallelizes expert routing across chips the way a multi-GPU CUDA rack would — everything is served off one chip's 819 GB/s memory bus. That's plenty of raw bandwidth for these token counts, but it means the extra headroom over the 256GB tier is almost entirely about which models fit at all, not about squeezing more tokens/sec out of ones that already fit elsewhere.

Apple M3 Ultra (256GB): the mid-memory Mac Studio configuration Apple introduced alongside the 96GB base and 512GB flagship in March 2025 — available as a BTO memory upgrade on either the base 28-core CPU (20P+8E)/60-core GPU die or the 32-core/80-core upgrade, at the identical 819 GB/s bandwidth either way, since M3 Ultra's UltraFusion-fused dual-die design doesn't bin bandwidth by CPU/GPU core count the way M3 Max bins by memory size.

With 248 GB of real headroom, this is the smallest Apple Silicon configuration this site tracks that fits a 350B+-parameter frontier MoE model at all: GLM 4.5 and GLM 4.6 (355B) run at Q4_K_M (245.60 GB, 9.6 tok/s). DeepSeek V3 and Kimi K2.6 still don't fit here — both need the 512GB tier. The MoE-vs-dense gap is stark at frontier scale: Llama 4 Maverick (400B, MoE) fits at Q3_K_M (220.00 GB) at 20.9 tok/s, while Llama 3.1 405B (405B, dense) needs the same Q3_K_M precision at almost the same footprint (222.92 GB) but manages only 3.3 tok/s — a 6x speed gap between two similarly-sized models, because MoE decode only reads its active parameters off the memory bus.

Same MLX and llama.cpp Metal maturity as the rest of the M3 Ultra lineup. At this file scale — GLM 4.5's Q4_K_M build alone is a quarter-terabyte download — a fast SSD and network connection matter as much as the chip; this is the one tier where local storage speed genuinely gates how quickly a new model gets from download to first token.

Apple M3 Ultra (96GB): the base Mac Studio configuration Apple introduced in March 2025, standard on the 28-core CPU (20P+8E)/60-core GPU die. Because M3 Ultra is built from two complete M3 Max dies fused with UltraFusion, this configuration shares its exact 88 GB of real headroom with the 96GB M3 Max — the difference is that the Max's 96GB build pairs with a binned, bandwidth-reduced die (300 GB/s), while the Ultra reaches 819 GB/s regardless of memory size: 2.73x faster at an identical usable capacity.

Because usable capacity matches the 96GB M3 Max exactly, this configuration fits the same 61 of 84 tracked models natively — what changes is speed, not the model list. This site's calculator projects Llama 3.3 70B at Q8_0 (86.35 GB) running 8.5 tok/s here versus 3.1 tok/s on the 96GB M3 Max, and Qwen 3.6 27B at BF16 (61.08 GB) at 12.0 tok/s versus 4.4 tok/s — both track the 2.73x bandwidth ratio almost exactly. MoE models benefit even more in absolute terms: GPT-OSS 120B reaches Q4_K_M (80.15 GB) at 61.5 tok/s here, genuinely real-time on a 117B-parameter model, versus 22.5 tok/s on the bandwidth-limited Max.

MLX and llama.cpp's Metal backend are both mature on this chip. Because usable capacity matches the 96GB M3 Max, the same macOS Metal gotcha applies: the default working-set limit (~75% of unified memory, about 72 GB here) sits below the 86.35 GB the Q8_0 Llama 3.3 70B fit above needs, so raising it with `sudo sysctl iogpu.wired_limit_mb=90112` (leaving 8 GB for macOS) is necessary before loading anything close to that ceiling — the exact same fix and figure as the 96GB M3 Max.