llama.cpp Metal gains M5 neural accelerator support, erasing MLX's prefill edge
MrPecunius · reddit · 2026-09-03
A Reddit user reports that llama.cpp's Metal backend appears to have gained support for the M5's matmul/neural accelerators, pushing GGUF prefill (e.g., Unsloth Q8) to 300-350 t/s on Apple M5 — on par with the best MLX results. Combined with GGUF+MTP already leading on token generation (19 t/s), the author sees no remaining reason to use MLX models, ending the need to switch formats based on prefill/generation mix. Tested with Qwen3 27B at Q8 on an M5 Pro; the post asks whether the community still sees a case for MLX or knows how to get mainstream MLX MTP working.
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18
- 12.3M tokens in 27 minutes for $0.14: agent session with 99.6% cache hit — teortaxesTex · 2026-10-18