Qwen3-Omni and Qwen3-ASR Audio Support Merged into llama.cpp

GitHub·medium signal

PR #19441 lands Qwen3-Omni-MOE (vision + audio) and Qwen3-ASR support in llama.cpp, with working GGUF checkpoints on HuggingFace. Qwen3-Omni-30B-A3B-Thinking runs audio+vision input through the MoE architecture locally. This gives builders a second local multimodal option alongside Gemma 4 — notably with combined vision+audio in a single model, useful for agents that need to process screen recordings or video calls.