Apple’s 2025 workstation built on the M3 Ultra (March 2025 announcement),
the first Mac with 512GB of unified memory. Three configurations: 96GB at
$3,999, 256GB at $5,499, 512GB at $7,499. All share 819 GB/s of memory
bandwidth and a 100W-class chip.
What it does well:
-
512GB in one box: at launch it was the only consumer machine that
could hold DeepSeek R1 or Kimi-class checkpoints whole, no offloading.
-
Efficiency: a full inference node at 100W of chip power - roughly
a tenth of a comparable multi-GPU server’s draw.
Where it falls short: MLX and GGUF runtimes, no CUDA; 819 GB/s is well
below the M4/M5 Ultra tiers, so tok/s on large models trails them. Apple
has superseded it twice since, and the M5 Ultra beats it on every spec.
Run it locally: LM Studio, Ollama, or mlx-lm. Check the
modeldex results for fit and estimated
tok/s per checkpoint on your memory tier.