has released Muse Spark Open, a massive 816-billion parameter open-weight model developed in partnership with . This natively multimodal Mixture-of-Experts system can interpret text and raw video sequences while utilizing a 1-million token context.
The model activates 42 billion parameters per inference and is optimized for the framework. Running on two Apple M5 Ultra Mac Studios via Thunderbolt, it hits 20 tokens per second, increasing to 36 tokens per second when scaling to four interconnected units.
This release targets high-end agentic and coding tasks, leveraging the 1.2 TB/s bandwidth of Apple's M5 Ultra silicon. By providing an open-weight model of this scale, Meta and Apple are enabling sophisticated local AI workflows on consumer-grade hardware.