Free AI Tool llama.cpp Just Got Faster on Your Phone

llama.cpp is a free, open-source program that lets you run AI chatbots directly on your own computer or phone, instead of paying a monthly fee to a cloud service. That matters because nothing you type gets sent to a company's servers — it all stays on your machine. The project just shipped a small update, b11005, aimed at a specific set of chips made by Qualcomm.

The bug it fixes is fussy but easy to picture. Before running a model, the software does a quick check to decide which parts of the model go where — some on the graphics chip, some on the main processor. One of those checks used a test input of all zeros, and the code mistakenly rejected it. That rejection pushed a piece of the model onto the slower main processor, which chopped the model's work into more separate chunks that couldn't run at the same time. On one Google model, the work got split five ways instead of two. Think of a relay race where the baton gets handed off three extra times — same runners, much slower finish.

The update simply accepts that all-zero test input, so the work stays together and the model runs more efficiently on Qualcomm hardware, including many Android phones and the new Snapdragon laptops. One more detail worth noticing: the fix was written with the help of an AI coding assistant, Claude Opus 5, which is credited in the notes. That's a small but telling sign of how software is increasingly being built.

The honest catch: this is plumbing, not a headline feature. If you use ChatGPT, Gemini, or Claude in a browser, absolutely nothing changes for you. Even llama.cpp users will only notice a difference if they're running models on Qualcomm chips — and the gain is modest, not dramatic. It's the kind of quiet maintenance work that keeps free, private AI usable without a cloud bill.