GitHub radar
MAGI-2 Preview: Open Video + Audio Generation by Sand.ai
Sand.ai releases MAGI-2 Preview, an open-source 114B-parameter audio-visual generation model capable of producing 10 second videos with synchronized audio given a text or image prompt. Local inference requires 8 NVIDIA Hopper GPUs.
MAGI-2 Preview is a multimodal audiovisual generative model developed by Sand.ai. It uses a shared architecture for both T2V and I2V (text to video/image to video with or without a text prompt and/or image reference) with 114 billion parameters, using a Mixture-of-Experts (MagiMoE) approach to activate only 6 billion parameters per token. It can generate 10 seconds of video with corresponding audio, given a textual prompt or textual prompt and image reference. It also does a preview denoising step before upsampling to 1080p resolution. You can download the model weights (307 GB) here on Hugging Face under the Apache 2.0 license.
Why a vibe-coder should care
If you’re interested in developing applications that take in text inputs and return short videos with corresponding audio, this is one of the first end-to-end open source models out there. However, note that this model needs 8 Hopper class GPUs, so it might be best suited for cloud deployments or research purposes.
▌ More finds