llama.cpp b9543 adds video support for Qwen3.5 and frame merge
New release lets you run Qwen3.5 video models locally with frame merging.
The latest b9543 release of llama.cpp, the popular C++ library for running large language models locally, brings significant multimodal improvements. The headline feature is experimental video support for the Qwen3.5 family of models, enabling inference on video inputs. Additionally, the release introduces 'frame merge' support for Qwen-VL-based models, which intelligently combines frames to reduce token count while preserving context, making video analysis more practical on consumer hardware.
The release also includes various clean-ups and design revisions for the LLaVA-UHD case. Build artifacts are provided for macOS (Apple Silicon with and without KleidiAI, Intel), Linux (x64/arm64 CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64/arm64 CPU, CUDA 12/13, Vulkan, HIP), Android arm64, and iOS XCFramework. This wide platform support ensures developers can run video-capable Qwen models on everything from laptops to edge devices, opening up new possibilities for local video summarization, captioning, and analysis without sending data to the cloud.
- b9543 adds video inference support for Qwen3.5 models directly in llama.cpp
- New 'frame merge' feature for Qwen-VL models reduces token usage by combining frames
- Builds include macOS, Linux (ARM/x64 with Vulkan/ROCm/OpenVINO), Windows (CPU/CUDA), and Android
Why It Matters
Enables local, private video AI on consumer hardware, reducing cloud dependency for multimodal tasks.