llama.cpp v0.8.0 Released with LiquidAI LFM3 and Audio Streaming
llama.cpp v0.8.0 introduces support for the LiquidAI LFM3 model family, DeepSeek V4.1 Flash, an experimental streaming-audio API, new server endpoints for real-time audio, and major backend performance enhancements. This major update expands local inference capabilities to modern multi-modal architectures and voice applications, strengthening llama.cpp's role as a foundational engine for local AI tooling. The release upgrades GGML to v0.28.0, adds a 50/50 SLRU cache for Mixture of Experts (MoE) models, and includes a new `/v1/realtime` endpoint implementing the OpenAI realtime protocol.
## BACKGROUND
llama.cpp is a widely used open-source software library designed to perform efficient inference on large language models using CPU and GPU hardware. It is co-developed alongside GGML, a general-purpose tensor library that provides the underlying tensor operations and the GGUF file format for quantized models.