Meta’s open weights reach Mac computers, local inference hits memory barrier

Meta launched its new model "Muse Spark Open", describing it as the most capable among open-weight models in the United States, with a notable feature that it is designed to run locally on Apple’s Mac computers.The massive model, which relies on 816 billion parameters and a context window of one million tokens, embeds advanced inference capabilities into desktop hardware.It targets complex programming tasks and self-operating agent workloads.

Evaluation results published by the company showed the model surpassing common open-weight models on the market. According to the launch data we examined, the model is designed as multi-modal, allowing it to understand code and handle media simultaneously. To demonstrate its field efficiency, the company integrated the model into its smart coding suite "Muse Code" and tested it in a full-agent workflow that included bug fixing, adding new features, and building the complete software distribution for the client application Mastodon.

Operational trials went beyond traditional code writing to reach complex creative production software. The company ran the model within the professional video-editing application DaVinci Resolve, where the agent automatically reviewed footage and cut scenes, illustrating the shift of models from text-chat interfaces to tools that directly control heavyweight desktop applications.

But this technical independence carries a steep hardware cost; the model requires, for local operation, massive compute capacity and at least one terabyte of RAM, as reported by Engadget. The data recommends running it on Mac Studio machines equipped with the M5 Ultra processor. While open-weight models give organizations flexibility, transparency and lower operating costs compared with closed models, limiting desktop operation to a terabyte of memory shifts the expense from cloud-inference bills to high-end hardware procurement budgets, prompting the company to make the model available through multiple API platforms as well.

This development imposes direct calculations on CTOs and development teams in the Gulf, Egypt and the Arab Levant.Financial institutions, creative production companies and coding centers that prefer open models to protect data privacy and trade secrets face a precise trade-off: either make upfront capital investment in high-end desktop workstations to run agents locally without leaking a single line of code outside their premises, or rely on cloud APIs with their associated data-transfer obligations. The model’s integration with programming and editing environments also shifts the developer’s required skill set from manually writing code to managing agents and directing automated build processes within the local work environment.

This step shows that the trajectory of open models is indeed moving from remote cloud servers to the developer’s desktop, but leveraging this power is no longer tied solely to obtaining a model license; it also depends on securing the physical memory required to accommodate its massive weights.