Meta released a large open-weight model on October 15 whose size puts it outside the range that fits on a single machine, and Apple, which has been at odds with the company for years, said the hardware to run it already exists.
The model, from Meta’s superintelligence research group, has 816 billion total parameters and activates about 42 billion of them per token, a design that keeps the computation per request far below what the total size suggests. It accepts a context of one million tokens and handles text, images, video and audio, with the emphasis placed on programming and long-running agent tasks.
Meta called it the strongest open-weight model produced in the United States and the first of its kind that can be run locally. Two Mac Studio machines, each configured with 512 gigabytes of unified memory and connected over Thunderbolt, pool about a terabyte of memory, which is enough to hold the weights.
The technical idea that makes this possible is a mixture of experts, in which only a fraction of the network is used for any given token. A model with hundreds of billions of weights can therefore run at the speed of a much smaller one, while requiring memory proportional to its full size. That combination, cheap computation and expensive memory, is what shaped the hardware answer in Cupertino rather than a data center.
Open weights change who can use a model rather than who can buy it. A company that downloads weights can run them without sending data to a vendor, fine-tune them on private material and keep the resulting system inside its own network. For organizations with contractual or legal reasons not to send material to a third party, that is a different product from an API.
The economics of local operation are unfavorable at the low end. Two workstations with that much memory cost more than a year of API calls for most users, and the electricity and maintenance add to the total. The case for running locally rests on volume, on privacy requirements or on the need to work without a network, and the buyers who satisfy those conditions are a specific subset of enterprises.
The political dimension is not incidental. Two companies that have spent five years arguing about privacy and platform rules appeared in the same announcement, with Meta’s chief executive thanking Apple’s in a public post and Apple’s chief executive describing the workstation as pushing the boundary of on-device computing. A shared technical interest in local inference gave both a reason to set the dispute aside.
That interest explains the odd alliance. A company selling local model execution benefits from more capable open models, and a company distributing open models benefits from hardware that can hold them. Neither gift costs the other anything, which is why a collaboration that would have been unlikely on policy questions is easy on silicon.
The competitive read is that open weights compress the price of capability. A model of this size available for download sets a floor under what a buyer will pay for a comparable proprietary system, and it gives developers a way to build without committing to a vendor’s terms. Labs that sell access respond by competing on serving quality, integration and support rather than on the model alone.
Support for four input types in one open model also matters for the applications being built. Agents that watch a screen, read a document and listen to a recording need all of those capabilities in one place, and stitching together separate specialized models has been the standard workaround. Combining them in one set of weights simplifies a deployment and removes a class of errors at the seams.
Analysts said the immediate effect will be visible in hardware demand rather than in model quality. Workstations configured with that much memory have been sold to developers and small research teams for a year, and a model that justifies the purchase gives those buyers a reason to add a second machine. The vendors that benefit are the ones selling memory rather than accelerators.
What the release does not answer is how the model behaves on the tasks it claims to lead, how many organizations will run it locally rather than through a hosted endpoint, and whether the license permits commercial use without restrictions that make it unworkable. Those details will determine whether an open release changes the market or simply raises the bar for the next one.
Licensing terms will decide much of that. A release that permits commercial use gives enterprises a path to run the model without a vendor relationship, while restrictions on scale or on competing products narrow the audience. Open-weight releases have historically varied widely on this point, and the terms matter more to adoption than the benchmark scores that accompany the announcement.
