llama.cpp Release b11941 Adds LiquidAI LFM3 Audio and Omni Support
Release b11941 of llama.cpp introduces support for LiquidAI's LFM3-Audio-2B and LFM3-Omni-2B models, bringing integrated text-to-speech voice options. Users can now leverage these new multimodal architectures directly within local runtime environments. This update expands local LLM runtimes beyond traditional text processing to handle advanced audio and omni modalities. Developers and privacy-conscious users can run cutting-edge voice models locally without relying on cloud-based APIs. The update includes configuration mapping for LFM3-Audio layout differences, disables flash attention for the LFM3-Audio depthformer, and leaves encoder convolutions and codecs unquantized to preserve quality. Voice selection is supported via the --tts-voice flag in llama-tts or the 'voice' parameter in the server /tts endpoint.
## BACKGROUND
llama.cpp is a popular open-source software library originally created by Georgi Gerganov that performs efficient inference on various large language models using the GGML tensor library. It allows users to run AI models locally on consumer hardware ranging from desktop GPUs to mobile devices.