davide-troiani/moeshift
View on GitHub ↗ ★ 95 C++
Memory efficient and high performance Mixture-of-Experts inference for Apple Silicon
4 mentions on Hacker News by 1 person: 2 stories, 2 comments · first 2026-11-03 · latest 2026-11-09
20082026
-
…For reproduction, I ran the benchmark using the stock benchmark runner included in the repo: https://github.com/davide-troiani/moeshift
comment by davidetroiani · 2026-11-09 · in M4 Pro vs. M5 Pro (24GB): Qwen 3.8 Flash Next at 374 Tok/S Prefill with MoEShift -
…much appreciated. Link to source code, architecture details, benchmarks and install instructions: https://github.com/davide-troiani/moeshift Cheers!
comment by davidetroiani · 2026-11-03 · in Show HN: MoEShift - Faster large MoE inference with half the RAM