Two Decades of Linux Expertise 400-9696-513 24/7 technical support

Meta open-sources the 816-billion-parameter model Muse Spark Open; two Mac Studios can run it fully locally

Meta open-sources the 816-billion-parameter model Muse Spark Open; two Mac Studios can run it fully locally

Meta has released Muse Spark Open, an open-weight multimodal large model with a total parameter count of 816 billion, described as one of the most capable open-weight models it has produced to date. The model uses a Mixture-of-Experts (MoE) architecture with 42 billion active parameters and supports an ultra-long context window of 1 million tokens. Its strengths are aimed at coding, multi-step reasoning, and AI agent tasks, with the main gains in long-horizon task execution, complex software engineering, and multi-step reasoning.

The most watched aspect of the release is the breakthrough in deployment thresholds: through the collaboration with Apple, Muse Spark Open can run fully locally and offline on two top-spec Mac Studio machines with M5 Ultra chips and 512GB of unified memory, using the latest MLX framework, with no cloud GPU cluster required. That means desktop-class hardware investment can now carry inference for frontier open models, pushing the local deployment path for ultra-large models further down.

For AI agent developers, a million-token context combined with open weights offers a new foundation for repository-level code understanding, long-document processing, and multi-step autonomous tasks; paired with local running capability, data-stays-in-premises private agent architectures are unblocked on both the cost and the compliance dimension.

For enterprise IT departments, this is worth folding into technology selection thinking. First, the capability frontier of open-weight models keeps rising, loosening the premise that nothing but a closed API will do. Second, the feasibility of running ultra-large models on workstation-class hardware with a unified memory architecture has been validated, opening up the design space for private deployment hardware plans. Local runnability, of course, is not production readiness — throughput, concurrency, and stability still need to be validated against real workloads, and an internal PoC is the sensible first step.

Source: Eastmoney

← Back to list Enquire about related services

Related reading

Call us Message
Home Services Consult Message