Compare

Speed on Apple silicon

How fast each engine reads your prompt and writes its answer, one request at a time: a 512-token prompt, a 256-token answer, every engine at 4-bit.

Etude runs on Pie, an open-source programmable serving system.

Higher is faster; the fastest in each row is in bold. doesn't fit in this Mac's memory, or the engine can't run the model. — not measured in this run. In Writing, “with speculative decoding” rows show the same model with drafting turned on: the engine guesses a few tokens ahead and checks them at once. Splash always drafts, so it has numbers only on those rows; MLX and llama.cpp show their usual numbers there, since drafting doesn't speed them up on these models.