llama.cpp Adds Support for Ling 3.0 VL Multimodal Model

llama.cpp has merged support for Ling 3.0 flash-VL, a multimodal model designed for native image and video understanding alongside text reasoning. The model features a 124B total parameter count with a sparse MoE architecture that activates only 5.5B parameters per token, optimizing inference efficiency while maintaining capacity.

Architecture and Capabilities

The model integrates visual information into agentic workflows through several key components:

• Visual Processing: A ViT visual encoder extracts features from images and videos, aligned with text representations via a two-layer MLP projector.
• Temporal Understanding: VideoRoPE encodes spatial and temporal positions, enabling tasks like event localization and long-video analysis.
• Long-Context Efficiency: A 42-layer hybrid backbone alternates KDA and Gated MLA layers at a 5:1 ratio, enabling efficient long-context processing across text, images, videos, and extended agent task histories, supporting a context window of up to 256K tokens.

This addition allows developers using llama.cpp to deploy Ling 3.0 VL for unified multimodal reasoning, planning, and verification tasks.