Show HN: Which Mac runs which LLM, and how fast

A web tool that predicts which Apple Silicon Macs can run which large language models and how fast. It reduces the decision to two determinants: whether a model’s weights fit into unified memory (parameter count × bits per weight plus overhead and KV cache) and the generation speed, which is set by memory bandwidth divided by weight size plus a fixed per-token overhead. Users pick a Mac, model, quantization, and context length; the calculator shows minimum RAM, weight size, and tokens/sec. Example: Llama 3.3 70B in Q4_K_M quantization uses ~42.8 GB of weights, needs ≈56 GB total (so a 64 GB M4 Pro is recommended) and yields about 5.7 tokens/sec. The site also offers rental pricing and measured benchmark links.

Underlying formulas are explicit: weight bytes = parameters × bits/weight/8, add 15% runtime overhead, KV cache, and 3 GB for macOS; speed = bandwidth/weight-size plus a fitted 3.3 ms overhead. The model was validated against eight dense runs on two machines (base M4 and M5 Pro), matching measured speeds to ~4.6%, and the page lists Apple chip bandwidths and memory tiers (e.g., M4 Pro 273 GB/s, M5 Ultra 1.2 TB/s). Important caveats are highlighted: the overhead is held constant across chips, many rows are extrapolated rather than measured, prompt-prefill is compute-bound (varies by model), MoE models require ~1.3× active weight reads, and binned chips with the same name can differ substantially in bandwidth.