The open‑source llama.cpp library released version b11761 on Oct 21, 2026. It corrects the NEON branch so the fp32‑to‑fp16 conversion now uses the hardware instruction instead of a software fallback. The change also adds a compile‑time error to catch similar macro mismatches. The fix applies to macOS Apple Silicon, Linux ARM64, and Android ARM devices listed in the release notes.
Why it matters
ARM‑based users will see faster model inference because the library now uses the processor’s native half‑precision conversion.
