Local AI speed on Mac and DGX Spark racks

Mac Studio M5 Ultra

Generation speed in tokens per second, one request at a time, 4-bit builds.

Tier Best model 4-bit size Smallest Mac that fits Price
Loading data…

Measured and estimated

Measured numbers are single-request generation speeds from public reviews and benchmark reports, without speculative decoding. Each one links to its source. Engines and quants differ between reports, so treat differences under about 15% as noise.

Estimates (striped bars, “≈ … est.”) fill cells where no run was found, and always start from a measured number. For the M1 Ultra, a measured M5 Ultra speed is scaled by the two chips' speeds in llama.cpp's Apple Silicon benchmark (83.73 vs 179.10 tokens/sec). For Nanbeige4.3-33B, a dense model, Qwen3.8 27B's measured speed on the same rack is scaled by file size. GLM-5.3 on one M5 Ultra is LLMCheck's estimate.

Racks of several M5 Ultras split the model across Macs over Thunderbolt 5 with RDMA (tensor parallelism in MLX). The anchor is Apple's measurement of Muse Spark Open (816B total, 42B active, 493 GB at 4-bit): 24.99 tokens/sec on 2 Macs and 36.73 on 4. GLM-5.3 has a similar shape, so 2× 256 GB is estimated at the same 25. Kimi K3.1 Pro on 4× 512 GB is the 4-Mac figure scaled by active parameters (42B / 104B). Kimi needs 7 Macs with 256 GB; no run that large is published.

Prices are US prices for the configuration shown (most memory, base storage), taken from Apple, NVIDIA or retailer listings in November 2026; a rack's price is the number of Macs times the price of one. The MacBook Pro price is approximate and the M1 Ultra, now discontinued, shows its 2022 launch price. Hover a price for its configuration and click it for the source.

Doesn't fit means the 4-bit build is larger than the rack's maximum memory. Sonnet 5.5 is the median API output speed from Artificial Analysis; it runs on Anthropic's servers, not a rack.

Data files: rack-speed-2026-11-18.csv, racks-2026-11-18.csv.