Free 'Run AI on Your Own Laptop' Tool Gets Key Bug Fixes

WHAT HAPPENED: A popular free software project called llama.cpp shipped a small maintenance update (version b10907) on September 11. llama.cpp is the engine that lets ordinary computers run AI chatbots locally — on a Mac, a Windows PC, an Android phone or even a small Linux box — without sending your conversations to a company's servers.

WHY IT MATTERS: This update squashes memory bugs affecting three families of AI models that are especially popular because they're free and open: DeepSeek, GLM and Cohere's models. The bug involved something called the 'KV cache,' which is essentially the AI's short-term memory of what you've said so far in a chat. When that memory was allocated incorrectly, these models could misfire or fail. Think of it like a waiter taking your order: if they scribble it on the wrong-sized notepad, parts of your order get lost. Same idea — now fixed.

THE SPECIFIC DETAIL: The patch also targets 'MTP,' or multi-token prediction — a trick where the AI guesses several words ahead at once to reply faster. That speed feature was misbehaving for these models, and the update adds automatic testing so the problem is caught earlier next time.

THE CATCH: This is plumbing, not a headline feature. If you don't already run local AI, nothing visibly changes for you today, and you'd never notice version numbers like b10907. But it's the kind of unglamorous upkeep that keeps the free, private alternative to paid AI subscriptions working. The release also includes ready-made downloads for macOS, Windows, Linux, Android and iOS, so tinkerers can grab it instantly.