Meta Launches Muse Spark Open to Rival Chinese Models

Meta has released Muse Spark Open, a massive open-weight multimodal model, positioning it as a direct competitor to leading Chinese open models like Qwen, GLM, and Kimi.

Meta has thrown its hat into the increasingly crowded open-weight model race with the release of Muse Spark Open, a sprawling 816-billion-parameter system built to handle text, images, video, and audio in a single unified model. The sheer scale of the release signals that Meta wants to compete not just on size but on versatility, aiming squarely at the kind of all-purpose capability that has made recent Chinese open models popular with developers. One of the model's headline features is its massive context window, which can stretch up to one million tokens. That kind of capacity allows the model to work with enormous amounts of information in a single session, whether that's long documents, extended video, or complex multimodal inputs that mix several media types at once. According to the report, Muse Spark Open performs roughly in line with top open-weight models from Chinese labs, including Qwen, GLM, and Kimi, across a range of benchmarks. That puts Meta in direct competition with a group of releases that have rapidly closed the gap with proprietary systems from Western AI labs over the past year, and suggests Meta sees open-weight releases as a way to stay relevant in that fast-moving corner of the industry. Notably, Muse Spark Open has also been optimized to run locally on Apple's M5 Ultra hardware, a detail that points to Meta's interest in making high-end multimodal AI accessible outside of cloud-dependent setups. By tuning the model for consumer-grade workstation hardware, Meta appears to be betting that developers and researchers will want powerful open models they can run on their own machines rather than relying solely on hosted inference.