Meta AI Research has published Muse Spark Open, an open weight, multimodal model it describes as the most capable United States open weight release to date. The model release, announced by Meta on October 15, 2026, publishes downloadable weights and documentation and highlights a deployment route that can run the model locally across two or more Apple Mac Studio machines using Apple’s MLX framework.
What Meta announced
Muse Spark Open is a mixture of experts architecture with 816 billion total parameters and 42 billion active parameters, and it supports very large context windows, which Meta lists at one million tokens. The company positions the model for coding, agentic and complex multi step reasoning tasks, and for multimodal inputs including text, image, audio and video.
Meta published the model weights and accompanying developer materials on its research and developer channels, and said the weights are available through major cloud model registries and third party API providers. Meta also published guidance and a set of recommended inference engines and kernels it has tested with the release.
Desktop deployment claim and the Apple connection
Unusually for a frontier class model, Meta emphasized a desktop deployment option. The company says Muse Spark Open can be quantized and deployed to run across paired Apple Mac Studio machines with Apple's M5 Ultra chips by using Apple’s MLX machine learning framework. Meta and Apple engineered the pathway so that, with sufficient unified memory across linked Macs, the model can be sharded and executed without requiring a large public GPU cluster.
Meta noted that a Mac Studio setup distributes the model’s memory demands across machines that expose unified memory to the GPU, and the published documentation includes measured token throughput numbers for reference configurations. Meta and Apple’s engineering cooperation on the MLX pathway is presented as a proof of concept for running highly capable models on high end desktop hardware.
Availability, licensing and ecosystem partners
Muse Spark Open’s weights and model files were posted to public model registries and are reachable through several API and hosting partners. Meta lists multiple third party platforms that offered live API endpoints on release day. The company also published a model license that places limits on certain commercial uses, and it says the open weight release underwent additional safety focused post training before distribution.
Meta provided a set of tested back ends and runtimes, including references to popular open source and commercial inference engines. The developer documentation encourages using hardware specific kernels and quantized formats for efficient deployment, and it explains tradeoffs between local desktop setups and cloud based hosting.
Why this matters and practical limits
The release matters for two reasons. First, it expands the set of highly capable models whose raw weights and deployment recipes are publicly distributed, enabling independent researchers, infrastructure vendors and application builders to inspect, adapt and run the model. Second, by publishing a tested route to run the model on linked high end Macs, Meta and Apple have demonstrated a desktop path that was previously uncommon for frontier class models.
That said, running Muse Spark Open locally remains resource intensive. Meta’s documentation and third party analyses make clear that the Mac Studio route requires large amounts of unified memory and careful configuration, effectively limiting practical local deployment to specialized workstations or institutional testbeds. For most developers and companies, cloud hosting or managed API endpoints will remain the more economical option.
Safety, evaluation and the broader context
Meta says Muse Spark Open received targeted safety post training and that it met release thresholds defined in its internal frameworks. Independent reviewers and benchmarking services have begun publishing evaluations, but the company cautions that comparative scores and real world performance vary by task, configuration, and kernel implementation.
The Muse Spark Open release follows a wave of open weight and research model releases this year, and it is likely to accelerate deployment and experimentation across academic labs, startups and infrastructure providers. At the same time, the release will intensify conversations about governance, licensing restrictions for commercial hosting, and operational safety for powerful, agentic models running outside tightly controlled data center environments.
Meta’s research and developer pages host the full technical write ups, runtime guidance and the Muse Model License. Third party providers published compatible API endpoints and additional deployment tools on release day, and independent benchmarks and hands on reports appeared in the days immediately following the announcement.
