Muse Spark’s open weights

On October 15, Meta released the weights of Muse Spark Open, a multimodal mixture-of-experts model built for coding, tool use and long-running agentic tasks. It has 816 billion total parameters but activates 42 billion per step, with a context window of up to one million tokens. Meta provides both BF16 weights and a more compact NVFP4 quantization.

The same day, Apple described running the NVFP4 build through MLX across multiple M5 Ultra Mac Studios linked over Thunderbolt 5. The model does not fit in one machine’s memory, so the baseline setup uses two computers. In Apple’s test, two Macs decoded at 15.74 tokens per second with pipeline parallelism and 24.99 with tensor parallelism; four machines reached 36.73 tokens per second.

Requirements for running it locally

Apple’s demonstrated configuration requires at least two Mac Studio computers because the model does not fit in one machine’s memory. The quoted speeds come from Apple’s tests. Local inference allows code and documents to be processed without sending them to a cloud API.

The license permits use, modification and distribution, but it is not unconditional. Products must display “Built with Muse Spark Open,” and Model-as-a-Service providers with revenue above $20 million need a separate agreement with Meta. Meta also recommends additional safeguards and human confirmation before an agent takes irreversible actions.

Sources