The 2025 Mac Studio generation, on M4 Max (36/64/96GB) and M4 Ultra
(192/512GB). Still on sale alongside the M5 refresh, and the price cuts
make the 512GB M4 Ultra the cheapest path to 512GB of unified memory at
$9,999 - the same memory Apple’s M5 Ultra 512GB charges a premium for.
What it does well:
-
M4 Ultra bandwidth: 1092 GB/s, second only to the M5 Ultra among
unified-memory machines on this site.
-
Memory ceiling: 192GB fits 70B dense models at 8-bit or 200B-class
MoE at 4-bit with headroom; 512GB runs the largest open weights.
-
Sustained inference acoustics: quiet under LLM load, desk-friendly.
Where it falls short: no CUDA; MLX-first runtimes, with GGUF as the
fallback. vLLM-class serving optimizations land later or not at all, and
the M4 Ultra’s bandwidth is ~9% below the M5 Ultra it replaced.
Run it locally: MLX (LM Studio, mlx-lm) and Ollama. The
modeldex results rank every seeded
checkpoint by fit and estimated tok/s on this exact config.