The llama.cpp project has released version b12425 with a change that allows it to read Qwen3-VL’s output-head size directly from the model tensor. The change supports Qwen3-ForcedAligner, whose output has 5,000 timestamp classes—fewer than the model’s vocabulary.
Why it matters
People running Qwen3-ForcedAligner with llama.cpp can use a model whose timestamp output is smaller than its vocabulary; the release notes do not specify setup requirements.