The release
Meta today released Muse Spark Open, which the company describes as the most capable U.S. open-weight model to date. The model carries 816 billion total parameters and 42 billion active parameters, paired with a 1-million-token context window. Meta states it is intended for coding and agentic tasks and has been optimized to run on the Mac Studio equipped with the M5 Ultra chip.
The announcement centers on a specific hardware pairing rather than a general cloud release. Two Mac Studio units, each fitted with an M5 Ultra and 512 GB of unified memory, run the model together under the MLX framework. This setup is presented as a complete, self-contained desktop system.
The collaboration
Meta worked directly with Apple to bring the model to the Mac platform. The result is a desktop configuration that uses two 512 GB Apple M5 Ultra Mac Studios running the MLX framework. Meta CEO Mark Zuckerberg publicly thanked Apple CEO John Ternus for the collaboration. Apple executive John Ternus responded that the setup allows a model of this scale to run efficiently from a desk and noted that the Mac Studio continues to advance on-device AI capabilities.
The exchange between the two executives was posted publicly on the day of release. Zuckerberg highlighted the efficiency of running a model of this size locally. Ternus pointed to the same configuration as evidence that the Mac Studio is extending the reach of on-device AI beyond smaller models.
Technical specifics
The model’s architecture mixes a large total parameter count with a much smaller active parameter set, a design choice that reduces compute demands while preserving capability for the target workloads. Meta positions the 1-million-token context window as a practical advantage for long codebases and multi-step agentic sequences. The company has not released additional benchmark numbers beyond its claim that the model excels at coding and agentic work.
The hardware arrangement is explicit: two M5 Ultra Mac Studios, each with 512 GB of unified memory, connected to run the model under the MLX framework. This configuration is presented as a complete on-premises solution rather than a cloud-dependent service. No other deployment targets were detailed in the announcement.
The active-parameter design keeps inference costs closer to those of a 42-billion-parameter model while still drawing on the full 816-billion-parameter store for accuracy on complex tasks. The 1-million-token window is sized to hold entire repositories or extended agent traces without chunking. Both features are presented as direct responses to the memory and bandwidth profile of the dual M5 Ultra setup.
Why it matters
The release shows Meta continuing to push large open-weight models while simultaneously courting Apple’s hardware ecosystem for local execution. For developers who already own or can acquire the specified Mac Studio setup, the model offers a concrete path to run an 816-billion-parameter system without sending data off-site. The public exchange between Zuckerberg and Ternus signals that both companies see value in making such models practical on high-end Apple silicon rather than restricting them to data-center clusters.
Whether this arrangement becomes a template for future large-model releases or remains a one-off demonstration will depend on how quickly Meta ships additional optimized variants and how Apple’s next hardware generations handle the memory and interconnect demands those variants will carry. The dual-Mac-Studio rig demonstrates that current unified-memory machines can host models previously assumed to require racks of GPUs, but only for users willing to buy and link two top-tier units. That narrows the audience today while widening the set of workloads that can stay on a single desk.
The absence of broader benchmark data leaves open the question of how Muse Spark Open compares with other open-weight systems on standard coding suites. What is clear is the hardware target: a self-contained, high-memory desktop pair running MLX. That choice prioritizes privacy and latency over easy scaling or multi-user access.
---
Sources:
No comments yet