llama.cpp Release b11156 Adds Support for Ling 3.0 VL Vision-Language Architecture
Open-source project llama.cpp released build b11156, adding support for the Ling 3.0 VL (Vision-Language) model architecture. The implementation integrates Ling 3.0 VL into the existing BailingMoeV3 architecture and manages mRoPE (Multimodal Rotary Position Embedding) sections. This enables local quantized inference for Ant Group's Ling 3.0 VL multimodal architecture across various consumer hardware platforms. It expands llama.cpp's support for open-weights vision-language Mixture-of-Experts (MoE) models. Technically, the Ling 3.0 VL implementation builds on the BailingMoeV3 architecture while keeping the shared normalization RoPE list intact during gating on mRoPE sections. Pre-compiled release binaries were published for multiple operating systems and backends, including CUDA 13, Vulkan, OpenVINO, and Snapdragon NPUs.
## BACKGROUND
llama.cpp is a popular C/C++ framework designed for efficient local execution of large language and vision-language models with minimal overhead. Ling 3.0 Flash VL is a multimodal Mixture-of-Experts (MoE) model developed by Ant Group that adds visual understanding capabilities to text language models. Multimodal Rotary Positional Embedding (mRoPE) is an attention positional encoding technique that extends traditional RoPE to encode multidimensional data like spatial image features alongside temporal text sequence tokens.