llama.cpp Ships Gemma 4 Audio and Qwen3 Omni Support — Multimodal Local Inference Moves From Experimental to Expected

GitHub·medium signal

llama.cpp has shipped support for Gemma 4 multimodal inference (text + image from day one, audio via E2B/E4B variants) and Qwen3 Omni models with both audio and vision input as GGUF models, in the same development window. Text and image processing are fully stable, while audio support is under active development with a 30-second maximum on Gemma 4 E2B/E4B. The convergence of audio and vision support in the dominant local inference engine signals that multimodal local inference is transitioning from 'experimental' to a baseline expectation for 2026 AI development.