Back to Articles
Authors: Arjun Reddy, Kesavan Sekar, Mahesh Kumar, Venkatesh Rajendran and Pritam ChristpinResearch attribution: osmAPI Research TeamRuntime repository: osmapi/SANA-WM-Bidirectional-on-Apple-SiliconBase model: Efficient-Large-Model/SANA-WM_bidirectional
SANA-WM is a 2.6B parameter controllable world model for image-to-video generation. The official release demonstrates a powerful idea: start from an image, provide a text prompt and a camera or action trajectory, and generate a coherent forward rollout of the world.
The public code and weights, however, arrived in a CUDA-first ecosystem. The model expected NVIDIA-oriented kernels, CUDA PyTorch, Triton, Flash Linear Attention, flash attention, xformers, and enough GPU memory to keep large pieces of the pipeline close together.
This article documents the Apple Silicon port: how we made SANA-WM Bidirectional run locally on an M3 Max MacBook Pro with 128 GB unified memory, while holding every stage below a strict 96 GB memory policy.
