U.S. tech giant Meta has released the open-weight model Muse Spark Open, a native multimodal model with a total of 816 billion parameters, of which 42 billion are active, and a context window supporting 1 million tokens. Designed specifically for programming and agent-based tasks, it excels at handling long-duration tasks, complex software engineering, and multi-step reasoning. Muse Spark Open is the first model of its kind capable of running on two Apple M5 Ultra Mac Studios using the latest MLX framework, with the combined performance equivalent to a supercomputer with 1TB of unified memory. Meta has completed testing of multiple local offline agent workflows based on this model. This release marks a rare public technical collaboration between Meta and Apple, who have long been at odds. Although Muse Spark Open's performance is slightly inferior to Meta's Muse Wonder, released earlier this week, Meta stated that it will not open-source its highest-performing model but will only make a slightly lower-performing version available to developers. In testing, Muse Spark Open outperformed competitors in certain projects but lagged behind Kimi K3 and GLM 5.3 in some professional capability assessments. Meta's previously launched Muse Wonder series models will be used to upgrade its personal AI agent, Muse, which was released on September 8 and has seen rapid download growth since launch, surpassing ChatGPT's initial mobile performance. Muse offers free and two tiers of paid services, representing a significant move by Meta to diversify its revenue streams beyond advertising through AI. Meta is making substantial investments in AI, having committed nearly $700 billion for AI-related infrastructure in the future. Despite year-over-year revenue growth in the second quarter, operating profit and earnings per share declined. Meanwhile, Meta disclosed its third-quarter revenue guidance. Goldman Sachs raised its target price for Meta, believing AI will drive advertising revenue growth, though the massive investments also pose cost pressures.