What Meta released on October 15
Meta’s announcement says it released the weights for Muse Spark Open on October 15, 2026. The model is multimodal and can take text, image, audio, and video input. Meta reports 816 billion total parameters, with 42 billion active per token, and a 1,048,576-token context window. Meta describes it as its most capable US open-weight model to date; that is the company’s description, not an independent comparison.
“42B active” does not mean a small download
A sparse mixture-of-experts model routes each token through some experts, but Meta’s developer docs say serving still loads all 816 billion weights into memory. Their listed BF16 checkpoint is about 1.65 TB; their NVFP4 checkpoint is about 537 GB. The docs estimate roughly 2 TB of VRAM for BF16, including serving overhead, and describe a four-GPU NVIDIA Blackwell node for NVFP4. Meta also says its MLX setup runs on two M5 Ultra Mac Studios. These are vendor configurations, not a promise that an ordinary laptop can host the model.
A practical check for a support-ticket assistant
- Illustrative example: imagine an assistant that labels fictional support tickets and drafts a reply. Write down the labels and what counts as an unsafe promise before comparing models.
- Build a small set of sanitized or made-up tickets: ordinary cases, missing order details, a refund request outside policy, and a message that should be escalated. Score whether the system follows the policy, asks when information is missing, and avoids inventing a resolution.
- Choose the deployment route before downloading weights. For self-hosting, check the exact checkpoint size plus memory for context, concurrency, and runtime overhead. A hosted inference provider avoids provisioning those GPUs, but has its own pricing, data handling, and service terms to inspect.
- Read the Muse Model License and its Usage Policy for your use case. The license requires a “Built with Muse Spark Open” notice in the product UI or associated documentation. If you and your affiliates exceed $20 million in aggregate gross revenue in any consecutive 12 months, it requires notice to Meta and a separate written agreement before offering the model or derivatives as a model-as-a-service. The clause excludes internal or research use and use as a component of your own non-MaaS product; check the full terms.
- Compare the candidate with your current system on the same test set. Record task success, failures, latency, and total operating cost; no single published model claim answers whether it fits your feature.
Weights available does not settle deployment fit
Meta’s launch page says its team tested agentic workflows and reports kernel and end-to-end speedups for particular hardware. This article did not download or run the model, reproduce those figures, or test a hosted provider. Check the license and provider terms for your planned use, then measure your own representative workload before deciding.