TL;DR: Meta released Muse Spark Open, an open-weights model that takes text, images, video and audio as input and reportedly runs on two high-memory desktops.
The Catalyst
Per DeepLearning.AI's The Batch, Muse Spark Open is Meta's return to open weights and is described as the strongest open-weights model from outside China, with native text, image, video and audio input. The post gives no parameter count, license, context limit or benchmark figures.
The Blast Radius
If the two-desktop claim holds, serving a multimodal model stops requiring a cluster and becomes a workstation purchase for local video/audio pipelines. Before planning around it, check the license terms and the exact memory footprint and quantization that claim assumes.
Source: x_com