Fix Makes Running AI on Your Own Computer More Trustworthy

Many people now run AI models directly on their own computers instead of using online services. This is faster and more private. The software that makes this possible, llama.cpp, just fixed a tricky bug that could cause models like Qwen3.8 to generate wrong text without warning. The problem only appeared on certain AMD and NVIDIA graphics cards when they used a technology called Vulkan. If you were affected, you might have gotten different answers each time you tried—or even invalid results while testing other features.

The bug was sneaky because it didn't crash or log an error. It happened deep in the software's optimizer, which reorders tasks to run faster. The optimizer wrongly assumed two tasks weren't connected when they actually used the same data. This caused the AI to produce "silently wrong tokens"—meaning you'd see a confident answer that was just incorrect. The fix teaches the optimizer to recognize these hidden connections, so it won't reorder things it shouldn't.

The update was released on August 28 and is already available. If you use llama.cpp with Vulkan, you should update to the latest version. For the rest of us, this is a good reminder: even the smartest AI relies on lots of careful engineering underneath. When that engineering is fixed, your AI becomes a little more reliable—and a little less likely to give you nonsense with a straight face.