The open‑source llama.cpp library now runs the softplus activation on Qualcomm Hexagon NPU, improving performance on supported Android devices.
NoteBy OossaPublished by Oossa: 1 min read
The llama.cpp project released version b11703 on Oct 17, 2026. It adds a vectorised softplus function that runs on Qualcomm’s Hexagon NPU (HVX). The new helper uses existing Hexagon primitives for exp and log, so the math stays the same as the CPU version. The change is listed under the Android arm64 build that supports Snapdragon CPUs, Adreno GPUs and Hexagon NPUs.
Developers can now run LLM inference faster on phones with Hexagon NPUs, extending llama.cpp’s low‑power use cases.
Models
Beijing's new draft limits exports of large AI models and 23 AI tools, but fast‑tracks open‑source releases.
Models
Gemini 4 Argon scored 97.5% on the toughest general‑problem‑solving test, second only to OpenAI's GPT‑6 Astra.
Models
The Linux kernel maintainer used an LLM to fix a cache‑contention bug, showing a 23‑times speed boost in synthetic tests.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
