llama.cpp Update Makes Free Local AI Faster on Intel Chips

The free software that lets you run AI chatbots on your own computer just got a speed boost — on Intel graphics chips. The project, called llama.cpp, released an update (labeled b11216) that rewrites one small but constantly repeated piece of math that AI models lean on. Think of it as swapping a hand-crank for a proper gear: same result, far less grinding. The math, a "Fast Walsh-Hadamard Transform," is how AI models shuffle and mix data internally so they can compress themselves and fit on smaller machines.

Why does a math shortcut matter to you? Because running AI locally is how a lot of people avoid $20-a-month chatbot subscriptions and keep private conversations off company servers. But local AI lives or dies on speed. Before this change, larger versions of that math step — the sizes real, capable models actually use — fell back to a brute-force method whose workload exploded: roughly double the size meant four times the work. The new version grows gently instead, so doubling the size means roughly double the work. Same answer, much less waiting.

Here is the honest catch. This is plumbing, not a shiny new feature. Nothing on your screen changes today, and the team could not fully test it on a real Intel GPU — they verified the math on a simulated CPU setup instead, and flagged that real-world testing is still needed. Only people with Intel graphics hardware benefit at all; NVIDIA and Apple users see nothing different.

Still, it is a reminder of how quietly free, local AI keeps improving. Thousands of small commits like this one are why running a capable chatbot on a laptop you already own has gone from a nerdy weekend project to a realistic alternative to paying monthly. The direction of travel is clear: the free option gets a little faster every few weeks.