Meta releases Muse Spark Open, runnable on Apple Mac Studios

Apple rewrote its MLX framework specifically so Muse Spark Open's 816 billion parameter weights can be split across two Mac Studios with M5 Ultra chips, pooling a full terabyte of unified memory (512GB per machine) and 1.2 TB/s of bandwidth per machine into what behaves like a single system^[1]. An Apple executive called it 'amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk'^[2]. The move signals that on-device inference for frontier-scale open models is now a joint engineering problem for chipmakers and model labs, not just a software afterthought, even if, as later sections cover, almost no individual actually owns the hardware needed to use it.

Meta bills Muse Spark Open as 'the most capable openly-available model built in the US to date'^[1], and outside coverage echoed that framing, calling it the most capable open-weight model to come out of a US lab so far^[3]. The scorecard is more mixed: Meta's own benchmarks show the model leading in Agentic, Knowledge and Reasoning, and Multimodal categories while trailing some Chinese rivals specifically on coding tests^[4], and head-to-head numbers put its Artificial Analysis Intelligence Index at 39 versus Xiaomi MiMo-V2.6-Pro's 46, and its Terminal-Bench 2.1 score at 76.4 versus Kimi K3's 85.0^[5]. That gap between the 'most capable US model' headline and a community reaction calling the model 'nerfed' or 'lobotomized' relative to Meta's own internal checkpoints is the real story here - Meta closed a hardware and accessibility gap with Apple before it closed the raw-capability gap with Chinese open-weight labs.

Meta's Chief AI Officer Alexandr Wang has said Meta will generally not open-source its most powerful models, which explains why the frontier Muse Wonder model launched days earlier remains proprietary while the comparatively weaker Muse Spark Open ships with downloadable weights^[6]. The split tracks the EU AI Act's Article 53, which exempts genuinely free and open models from certain documentation obligations only until systemic risk is present, described directly as 'the regulatory structure that incentivizes companies like Meta to release weaker models openly while keeping frontier models closed'^[6]. In other words, Muse Spark Open's openness is partly a regulatory-compliance choice, not purely a generosity one - it is calibrated to stay just under the systemic-risk line that would otherwise keep it closed too.

Even with Apple's MLX work, genuinely local deployment remains out of reach for almost anyone outside an enterprise budget - a four-unit, 256GB M5 Ultra Mac Studio configuration runs roughly $43,196 before tax and accessories^[5], and even a compressed 4-bit version of the model still occupies about 493GB, before accounting for the extra memory needed to actually run it^[7]. The Muse Model License adds its own friction for anyone who clears the hardware bar - companies generating more than $20 million in gross revenue over any trailing 12-month period must sign a separate agreement with Meta before offering the model as a hosted API service, and any product built on it must be labeled 'Built with Muse Spark Open'^[4]. Commentators have flagged the revenue-threshold clause specifically as the kind of fine print that makes the release less 'open' than a true Apache 2.0 license^[8].