The open‑source llama.cpp library released version b12385 on November 14, 2026. The update addresses a bug in the Qwen2VL model where scaling factors for feed‑forward network layers were not applied. As a result the model produced nonsensical outputs, with perplexity around 1 e6 on CUDA and NaN on CPU. The fix passes the correct scales to the build routine, restoring normal operation.
Why it matters
Developers using llama.cpp can now run Qwen2VL reliably, avoiding broken results in their applications.
