Elix speeds local Mac inference up to 28%

MacPaw’s in-development Elix runtime reports up to 28% faster local inference than stock MLX, moving an LFM2.5-1.2B 4-bit model from 406 to 521 tokens per second on an M3 Ultra and from 222 to 285 on an M4 Pro. The live Elix benchmarks span four Apple chips and arrive as MacPaw partners with Liquid AI on local models and memory, according to TechCrunch’s August 5 report. For Mac developers evaluating the economics of running AI on Apple hardware, the useful signal is a Swift-first stack with zero per-query cloud cost; the caveat is that Elix remains in development and its performance figures are vendor-run.