New Update Speeds Up AI on Qualcomm Phones

Think of llama.cpp as a free engine that lets AI models run directly on your phone or computer — no cloud servers required. This week, its developers released an update (b10687) with targeted fixes for Qualcomm's Adreno GPUs, the graphics and AI chips found in many Android phones. They tweaked how the software handles certain math gears inside the chip, so common AI models like gpt-oss-20b and gemma-3n get from zero to first response significantly faster.

Why does that matter to you? If you're using apps that run AI on your device — like a smart keyboard, photo editor, or local chat assistant — this can feel like the difference between waiting and getting an answer almost instantly. On one high-end chip, the Adreno X2-90, the gpt-oss-20b model gained a 25% speedup during the "thinking" phase (called prefill). On an older chip, the Adreno 740, another model got 9% faster. These may sound like benchmark numbers, but they translate directly to less thumb-twiddling.

The catch: these gains only apply to specific GPU families and certain models. The update does nothing for phones with other chips, and it doesn't speed up the later "writing" phase of AI output. Also, the improvements are for developers and tech-savvy users first — you won't get them automatically unless the apps you use adopt the updated code.

Still, this is a big step toward making on-device AI a real everyday experience. Faster, private AI with no internet requirement isn't just a trick — it's becoming the norm, and updates like this make it much more pleasant. For anyone tired of apps that need to phone home before answering, this is good news.