Meta released the weights for Muse Spark Open on October 15. The model has 816 billion total parameters, with 42 billion active on each token, and a one-million-token context window. It takes text, images, video, and audio as input and returns text. Meta worked with Apple so the model runs on two or four M5 Ultra Mac Studios through Apple’s MLX framework. The machines link over Thunderbolt 5 with remote direct memory access. When run locally, prompts and intermediate results stay on the hardware. Meta has no view of them.

Llama let developers download the weights, fine-tune them, and run them on their own infrastructure for years. Muse Spark Open extends that record into larger agentic work. Meta says it is the most capable U.S. open-weight model to date. The company demonstrated the model shipping actual iOS changes to the open-source Mastodon client Ice Cubes. Those changes included bug fixes, new features, and a full build. In a second test the model drove DaVinci Resolve end to end, reviewing footage frame by frame to select the strongest moments and crop them to vertical format. Both demonstrations ran entirely on the linked Mac Studios.

Two or four Mac Studios hold the model

Apple’s own tests show the quantized NVFP4 version needs roughly 493 GB for the weights alone. Intermediate computations push the total higher. A single 512 GB Mac Studio is not enough. Two machines are the minimum. Four machines with 256 GB each also work. The 512 GB M5 Ultra configurations were not widely available at launch and are expected later in October. M5 Ultra Mac Studios start above $5,000. Configurations with 256 GB of memory already sit well above $10,000. A two- or four-machine cluster costs tens of thousands of dollars before networking and power. Most teams will still use the cloud APIs that Meta also offers through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights are also available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

Apple published performance numbers the same day. At an 8,192-token context the two-machine pipeline setup produced 15.74 tokens per second. Tensor parallelization on the same two machines raised that to 24.99 tokens per second. Four machines with tensor parallelization reached 36.73 tokens per second. Time to first token dropped from about 8.4 seconds on the two-machine pipeline to 4.8 seconds on four machines. Those figures measure generation after the context is loaded. Longer contexts increase the wait. The numbers come from Apple’s MLX benchmarks and have not been independently reproduced at the same scale.

The license requires attribution and a revenue notice

Meta released the weights under the Muse Model License dated October 15, 2026. The license grants free rights to use, copy, modify, fine-tune, distill, merge, publish, distribute, and sell copies of the model, subject to three conditions. First, every product or service that uses the model or a derivative must prominently display “Built with Muse Spark Open” in the interface, documentation, website, or settings menu. A derivative includes any model trained or improved using the original model or its outputs. Second, a company that offers artificial intelligence models as a service via API or similar interface, and whose aggregate gross revenue with affiliates exceeds $20 million in any consecutive 12-month period, must notify Meta and enter a separate written agreement before offering Muse Spark Open or a derivative as part of that service. Internal use, research, and using the model to power features of the company’s own product are exempt. Third, use must comply with the attached Usage Policy, which restricts certain high-risk applications. Meta claims no ownership of the model’s outputs.

Those terms matter for enterprises. A company that runs the model on its own servers for internal coding agents faces the attribution requirement but not the $20 million clause. A cloud provider or platform that resells access above the revenue threshold needs a commercial agreement. Distilled or fine-tuned models carry the same obligations. The license is more permissive than some earlier Llama terms on commercial scale, yet it is still a custom contract rather than a pure open-source grant. Teams evaluating the model should read the full text on the Hugging Face model page before they budget time or money.

CIOs can test local or cloud paths

CIOs evaluating coding agents should ask three questions. Can the model stay on machines the company already owns, or does the useful path still require Meta’s cloud or a multi-Mac Studio cluster? What does the Muse Model License actually allow for commercial use, redistribution, and distillation? How will Meta support the weights and the MLX path six months from now? The answers decide whether this is a practical option or a one-week demonstration. Hardware cost and power draw for a two- or four-machine cluster are real. Cloud APIs avoid the capital outlay but reintroduce the data-residency and vendor-visibility questions that local execution was meant to solve.

Open weights lower the cost of inference once the model is downloaded, but they raise the cost of evaluation, safety review, and operational support. A company that runs Muse Spark Open on its own Macs or servers owns the prompts and the outputs. It also owns the responsibility for monitoring behavior, applying guardrails, and deciding when an update is safe. Meta’s additional safety post-training and its Superintelligence Scaling Framework evaluation are starting points, not a substitute for the buyer’s own tests. One incident in a production coding agent would travel farther than the same incident in a closed API, because the weights are already in the field.

Apple gains a high-profile demonstration of its unified-memory architecture and MLX framework. The company has not announced any plan to incorporate Muse Spark Open into Apple Intelligence or Siri. The collaboration is limited to the developer tooling path. For Meta, the public support from Apple’s CEO is useful marketing. For developers who already run open models on Apple Silicon, the MLX adapter and the published benchmarks reduce the time required to stand up a test. Neither side has said the work will continue beyond this release.

The practical next step for most organizations is a controlled pilot. Download the NVFP4 weights, stand up a two-machine cluster if the capital is available, or use one of the day-zero cloud endpoints. Measure token cost, latency at realistic context lengths, and success rate on the coding tasks the team actually runs. Compare those numbers with the closed Muse Wonder models that Meta released on October 12 and with the open models already in production. The pilot will show whether local execution changes the economics or the risk profile enough to matter.