Meta released Muse Spark Open on October 15, 2026, giving developers downloadable weights alongside hosted access. Its announcement says the model can run on two Apple M5 Ultra Mac Studios using MLX. That is a specific hardware configuration, not a promise that ordinary desktops can run it. Source: Meta's release announcement.
What the model offers
Muse Spark Open targets coding and agentic tasks. Meta's model card describes text, image, video and audio inputs, with text output and a roughly one-million-token context window. Those capabilities make it a candidate for workflows involving repositories, screenshots and documents, but suitability still needs testing against a team's own tasks. Source: Meta's model card.
Active parameters do not determine memory needs
Meta documents 816 billion total parameters and 42 billion active parameters. The sparse architecture activates only part of the model per token, but serving still requires loading the full weights. Its BF16 checkpoint is about 1.65 TB; the NVFP4 version is about 537 GB. Runtime overhead, activations and the key-value cache require additional memory. Source: Meta's deployment documentation.
The operational implication is straightforward: do not size a deployment from the active-parameter number alone. Test the intended context length and concurrency, not merely whether a checkpoint loads. A single successful demonstration is not a capacity plan.
Downloadable does not mean unrestricted
The release uses the custom Muse Model License. It requires attribution and includes a separate-agreement condition for qualifying Model as a Service businesses whose aggregate revenue exceeds $20 million over a consecutive twelve-month period. The license states exceptions for internal use and certain product features. Review the exact terms for the intended business model rather than treating the weights as an unrestricted release. Source: Muse Model License.
Choose the deployment boundary before the infrastructure
Meta documents both self-hosting and hosted-provider access. Our assessment: hosted inference is a sensible first comparison when the immediate question is output quality; self-hosting merits evaluation when infrastructure ownership is itself a requirement. Neither route establishes security, compliance or cost savings automatically. Source: deployment options.
For an agent pilot, define allowed tools, approval points and recovery procedures before enabling consequential actions. Meta's model card recommends additional safeguards tailored to the use case, including risks from external-tool integration. Source: deployment safety guidance.
The useful takeaway is greater deployment choice, not a universal replacement for hosted models. Begin with a bounded workflow and compare task success, latency, memory demand and operating effort before committing to hardware.