llama.cpp Adds Support for Ling 3.0 Flash VL Multimodal MoE Model
A new pull request (#29151) to llama.cpp by developer aetherbird adds support for Ling 3.0 Flash VL. The model is a 124B-parameter sparse Mixture-of-Experts (MoE) vision-language model with 5.5B active parameters per token and a 256K token context window. Bringing native video and image processing capabilities of Ling 3.0 Flash VL to llama.cpp enables efficient local execution of large multimodal models on consumer-grade hardware. This expands open-source capabilities for complex tasks like long-video analysis, visual reasoning, and agentic workflows without relying on cloud APIs. The architecture features a 42-layer hybrid backbone alternating KDA and Gated MLA layers in a 5:1 ratio, alongside VideoRoPE for spatial and temporal encoding in videos. It integrates visual inputs via a ViT encoder paired with a 2-layer MLP projector to align image and text representations.
## BACKGROUND
Multi-Head Latent Attention (MLA) and its extensions like Gated MLA compress Key-Value (KV) cache requirements, making long-context processing much more memory-efficient. Additionally, VideoRoPE adapts rotary position embeddings to handle both 3D spatial features and temporal ordering, allowing multimodal models to track continuous changes over time in video frames.