According to The Verge, Meta unveiled an open-weight AI model named Muse Spark Open. The company calls it the "most capable US open-weight model to date." It is also said to be the first in its class capable of running on a desktop, specifically with two Apple M5 Ultra Mac Studios and the latest MLX framework. Rather than making a generic cloud deployment claim, the announcement ties a large model release to a named hardware setup.
With 816 billion total parameters, the model is described as natively multimodal. Meta claims it performs well on long-horizon tasks, complex software engineering, and multi-step reasoning. These are Meta's own capability characterizations, not findings from an independent evaluation. Benchmark numbers, evaluation methodology, and a model card are not included in the supplied material.
The collaboration is the notable element in this situation. Although Meta and Apple are typically rivals, The Verge reports that the two companies coordinated on the launch and are now offering each other open endorsements. A post on X from Meta CEO Mark Zuckerberg announced that the weights would be released that same day, and he gave credit to @johnternus and Apple for the model's performance on the new Mac Studios. Congratulations to the team came in a reply from Apple CEO John Ternus, who said that running a model of Muse Spark Open's scale and caliber so efficiently right from a desk is remarkable, and he added that Mac Studio keeps broadening what on-device AI is capable of.
Alexandr Wang, Meta's chief AI officer, told Axios that the two companies had spent several months working together, and The Verge relays that reporting. In the same interview he joked that he was glad to see the two parent companies getting along again. Axios's reporting, as passed along by The Verge, is where the multi-month timeline comes from. Apart from the launch and the MLX work, what that collaboration entailed is not spelled out in the supplied evidence.
Apple also published a new technical post covering MLX, described as an array framework for machine learning tasks on Apple Silicon, along with Muse Spark Open and two other open models. That post is the documented technical artifact behind the performance claim, but the evidence here does not include its contents, so the specific optimizations, supported MLX version, or measured throughput remain unknown.
For developers, the practical shape of the announcement is that a very large open-weight model is being positioned as runnable on local Apple hardware rather than only in a datacenter. The claim is tied to a specific configuration: two M5 Ultra Mac Studios plus the latest MLX framework. Nothing in the evidence states memory requirements, quantization, token throughput, latency, or whether smaller single-machine setups are viable.
That specificity matters for anyone budgeting a local inference rig. A two-machine configuration implies a meaningful hardware outlay and some setup work to distribute the model across machines, and the evidence does not say how that distribution is handled or whether it requires custom tooling beyond MLX. Treat the desktop claim as a vendor and partner statement about one configuration, not a general statement about laptop-class machines.
The open-weight framing is also narrower than it may sound. Meta calls it an open-weight model, and Zuckerberg said the weights were releasing on announcement day, but the evidence does not describe the license terms, whether commercial use is permitted, or any usage restrictions. Freelancers and small studios should verify licensing before building on it.
There is a competitive subtext worth noting without overreading it. Meta's "most capable US open-weight model" phrasing is a comparative claim made by the releasing company, and the evidence supplies no third-party comparison or evaluation to support it. The Apple partnership is presented as a distribution and performance story for Apple Silicon, not as an endorsement of the model's quality.
The timing of the two companies' public statements is documented, but the evidence does not establish any broader coordination beyond this launch. The Verge notes the companies have been collaborating for months per Axios, which is a claim attributed to Meta's chief AI officer rather than independently verified in the supplied material.
What remains genuinely unknown is substantial. There is no pricing information because the model is open-weight, but there is also no information on hosting costs, no benchmark table, no context window figure, no training data description, and no independent reproduction of the two-Mac-Studio claim. The evidence also does not say whether the model is available through any hosted API.
For this audience, the reasonable takeaway is procedural: a large open-weight multimodal model now has a documented Apple Silicon path, which is relevant if you already own or plan to buy Mac Studio hardware and want local inference for long-horizon coding or multi-step reasoning work. Before committing, check the actual license, the MLX post's stated requirements, and whether the two-machine setup fits your workflow.
The announcement is best read as a launch with a strong partner story and thin independent verification. The concrete, attributable facts are the model name, the 816 billion parameter count, the natively multimodal description, the two-M5-Ultra-Mac-Studio configuration with MLX, the public statements from Zuckerberg and Ternus, Wang's comment to Axios, and Apple's technical post. Everything about real-world performance, cost, and licensing terms still needs to be confirmed from primary documentation.