Free 'Run AI on Your Own Laptop' Tool Just Fixed a Math Bug

Llama.cpp is the free, open-source engine behind a lot of "run AI on your own device" apps. Instead of sending your questions to a company's servers, it runs the AI right on your laptop, desktop or phone — no subscription, no monthly bill, and your data never leaves your machine. On September 16, the project released a new version called b11001.

The headline fix is a small but nasty coding mistake. Someone had copied a chunk of code meant for 32-bit numbers and pasted it into a section handling 16-bit numbers. The chip then read twice as many bytes per number as it should have. The result: for a 4x4 grid of 16-bit values, 14 of the 16 came out in the wrong place, producing garbage instead of the correct answer. It's the programming equivalent of grabbing the wrong measuring cup and doubling every ingredient.

Who actually felt this? Only devices built on SpacemiT's RISC-V chips — a newer, low-cost chip family found mostly in hobbyist boards and some budget machines. If you use a normal Intel, AMD or Apple laptop, or a typical Android phone, nothing changed for you. That's worth saying plainly, because bug-fix notes like this can sound scarier than they are.

The bigger story is the rest of the release. Alongside the fix, the project published ready-made downloads for Windows, Mac, Linux, Android and iOS, with support for NVIDIA, AMD and Intel graphics cards. That matters to non-engineers: "download it, double-click it, done" is exactly what lets ordinary people run a private AI assistant without handing their documents to a tech giant. Free, local AI keeps getting easier to set up — and this update keeps it accurate.