Two companies that have spent a decade at each other's throats just shipped an AI model together. Meta released Muse Spark Open, an open-weight system it calls the most capable ever put out by a US company, and in the same breath announced it had worked with Apple to make the thing run smoothly on the Mac Studio. Mark Zuckerberg and Apple's new chief executive John Ternus traded congratulations on X. Meta's AI chief Alexandr Wang described it, with a straight enough face, as "mommy and daddy playing nice again."
The détente is the surprise, because the history is ugly. Apple built its brand on privacy; Meta built its fortune on tracking. When Apple let iPhone owners opt out of app tracking in 2021, it cost Meta an estimated 10 billion dollars a year. So a joint AI launch reads as less a friendship than a division of labour. Meta wants to own open models. Apple wants to sell the hardware that runs them. Muse Spark Open, with 42 billion active parameters and a million-token context window, is tuned to run on Apple silicon, specifically the M5 Mac Studio Ultra. The catch for anyone wanting to try it at home: you need two of them, and the base machines start at 5,499 dollars each.
Strip away the novelty of the pairing and a sharper motive appears. The open-weight race is one the United States is losing, and both companies know it. On OpenRouter's leaderboard, four of the top five models last week were Chinese. On Artificial Analysis's Intelligence Index, Muse Spark Open scored a respectable 39, ahead of several Western open models, yet still behind Z.ai's GLM 5.5 and Moonshot's Kimi K3. A string of recent US open releases, from Reflection, Nvidia, Thinking Machines and Mistral, have all landed behind their Chinese counterparts. As the researcher Nathan Lambert put it bluntly: some of the best-funded researchers in America may simply not win, because "the Chinese are very very good at building LLMs."
Why does an open model matter enough for Meta to swallow its pride and call Apple? Because open weights can be downloaded and run on your own machines, which is both a security pitch and a hardware pitch. A company running a model on its own Mac Studios never sends data to anyone's cloud, an argument that lands harder each time an AI agent is caught breaking into a system it should not have touched. It also sidesteps the data-centre and chip bottleneck that has defined the frontier labs. On-device AI is the thread that ties Meta's software ambition to Apple's silicon, and it is why these two found common ground at all.
There is a competitive subplot too. Days before Spark Open, Meta showed off Muse Wonder, a closed flagship it claims beats OpenAI's GPT-Astra, Anthropic's Fable 5 and Google's Gemini 4 on the same index. Taken together, the two launches look like Meta trying to re-enter a US frontier conversation that Anthropic and OpenAI have dominated. Whether an open model that still trails Beijing's best counts as catching up, or just catching the eye, is the open question. For now, the most interesting thing Meta built this week may not be the model. It is the truce.