Meta Releases Muse Spark Open: 816B Open-Weight AI Model Runs Across Apple Mac Studios
Meta has released Muse Spark Open, an 816B-parameter open-weight model for agents, with Apple-backed local inference across M5 Ultra Mac Studios.
Summary
Meta released Muse Spark Open on October 15, 2026, publishing the weights of an 816-billion-parameter, natively multimodal model aimed at coding and long-running AI agents. The release brings a frontier-class Muse model to downloadable, customizable deployment rather than access exclusively through a vendor-managed interface. It is also available from several hosted inference providers.
With Apple, Meta demonstrates local inference across two M5 Ultra Mac Studios using MLX; Apple measures faster generation across four machines. This is not a model for an ordinary laptop.
Quick Take
- New open weights: Muse Spark Open ships with 816B total parameters, 42B active parameters and a 1-million-token context window.
- Multimodal in, text out: The official model card lists text, images, video and audio as inputs, with text output.
- Two-Mac local setup: Meta demonstrates inference across two M5 Ultra Mac Studios; Apple reports 36.73 tokens/second on four at an 8,192-token input context.
- Download or host: Weights are available through Meta's Hugging Face collection; several third-party API platforms offer hosted access.
- Not unrestricted open source: The Muse Model License includes attribution, prohibited uses and a special condition for large model-as-a-service businesses.
What launched—and what did not
Meta calls Muse Spark Open its strongest U.S.-developed open-weight model to date. That is Meta's characterization, not an independently established ranking across every open model. The release has 816B total parameters but activates 42B for a given token through a mixture-of-experts architecture. A large total parameter count therefore should not be mistaken for dense-model compute per token.
The model card confirms 1M-token context, multimodal input, text output and configurable reasoning effort. Its coding and agent benchmarks are vendor-reported, not AI World Scope testing.
Unlike this week's managed Muse Wonder releases, Muse Spark Open provides downloadable weights. Self-hosted inference has infrastructure costs rather than a fixed vendor token rate; hosted prices vary.
Original analysis: three ways to deploy the model
| Deployment path | What you control | Main practical constraint |
|---|---|---|
| Download weights | Checkpoint, infrastructure, custom tuning and local data path | Substantial memory, engineering and license compliance |
| Two M5 Ultra Mac Studios | On-premises inference through MLX and connected machines | Two high-memory desktops; not a single consumer Mac |
| Hosted third-party API | Application integration and provider selection | Provider pricing, privacy terms, uptime and data handling |
Control has a cost: open weights reduce endpoint dependence but require hardware, evaluations and application-level safety controls. Benchmark real workloads first.
Original analysis: what Apple's numbers actually show
Apple estimates 493GB for NVFP4 weights before runtime overhead; one Mac Studio cannot fit the model. Meta's card recommends at least 700GB of GPU memory for conventional NVFP4 deployment—different assumptions, not conflicting figures.
Apple's reported 36.73 tokens/second comes from four M5 Ultra Mac Studios with tensor parallelism at an 8,192-token prompt context and a 128-token generation test. A two-machine tensor-parallel setup measured 24.99 tokens/second under the same listed conditions. These are Apple measurements—not guaranteed throughput for long-context, real-world agent workloads.
Buyers should compare latency, hardware, privacy and maintenance costs with hosted APIs.
License and safety limits
Meta's Muse Model License permits downloading, adapting and many commercial uses, but it is not an unrestricted open-source license. Products using the model must display a “Built with Muse Spark Open” notice. Businesses offering models as a service with more than $20 million in aggregate gross revenue over a consecutive 12-month period must notify Meta and secure a separate written agreement before offering the model through that service. The clause has explicit exceptions, including certain internal and product-feature uses.
Meta reports additional safety training. Independent testing, tool permissions and human approval for irreversible actions remain necessary.
What to Watch Next
Watch independent agent evaluations, actual deployment costs, hosted pricing and license changes.
AI World Scope Take
NEW, not an update: the weights, custom license and Apple-backed local inference work make this a materially different release from Meta's earlier Muse Wonder launch. The breakthrough is not that an 816B model runs on a typical desktop; it is that a high-end, multi-machine desktop cluster can now serve a frontier-class open-weight agent model. For most developers, hosted access remains the more realistic first test.
Sources & Documentation
Sources used for this article, with source type and publisher shown where available.
- officialIntroducing Muse Spark Open, the Most Capable US Open-Weight Model to DateVisit Source
- documentationMuse Spark Open official model card and Muse Model LicenseVisit Source
- officialLocal and Distributed Large LLM Inference with MLX and RDMA on Mac StudioVisit Source
- newsMeta releases open-weight Muse Spark model for coding, agentic tasksVisit Source
- newsMeta launches an open-weight AI model that runs on MacsVisit Source