llama.cpp Release b11488 Adds Support for Liquid AI's d1-omni-600M Model
llama.cpp release b11488 introduces support for Liquid AI's d1-omni-600M decision model architecture. The release also includes server updates for multimodal audio input handling and CUDA fixes for conformer GLU sigmoid operations. Supporting Liquid AI's d1-omni decision model enables fast, non-generative classification and decision-making on local devices without relying on token generation. It expands llama.cpp beyond traditional auto-regressive language models into real-time multimodal decision tasks. Technical updates include renaming the internal decision type to 'lfm2-d1-omni', making images an alias for files in the server API, and scoping memory-less LFM2 to non-causal attention. Pre-built binaries were also updated across supported systems, including Linux, macOS, Windows, and Android.
## BACKGROUND
llama.cpp is a popular open-source inference engine designed for high-performance execution of language models locally using C/C++. Liquid AI's d1 decision models are multimodal architectures designed to output direct probabilities for specific options in a single forward pass rather than generating text step-by-step.