Meta and Apple, two rivals that for years avoided or publicly attacked each other, are now sitting side by side in a partnership that shifts the balance of power in artificial intelligence. The announcement came last Thursday (October 15), when Meta released the weights for Muse Spark Open, a language model with 816 billion parameters capable of running locally on two Mac Studios with M5 Ultra chips — ironically, hardware that Apple itself manufactures and hasn't yet put on sale.
The model is the most capable open-weight model ever built by a U.S. lab. With 816 billion total parameters but only 42 billion activated per token (a mixture-of-experts architecture), it processes text, images, audio, and video with a context window of one million tokens. In practical terms, it can read an entire book while simultaneously analyzing attached images and audio clips without losing track.
The choice of a sparse mixture-of-experts architecture is not accidental. While dense models from companies like OpenAI and Google activate all their parameters on every call — meaning inference costs scale linearly — Muse Spark Open activates only about 5% of its parameters per token. The result is a model that, in terms of operational cost per call, competes with neural networks a fraction of its size while maintaining frontier-level reasoning capability.
The hardware detail, however, is the bottleneck. Two Mac Studios with M5 Ultra chips and 512 GB of unified memory each offer a unified terabyte — the minimum required to run the 4-bit NVFP4 checkpoint (roughly 530 GB in practice). Apple's MLX framework, developed internally to make it easier to run AI models on Apple silicon, splits every layer across both machines, with communication via Thunderbolt 5 and RDMA. Four 256 GB Mac Studios also work, but the total hardware cost exceeds $40,000 — a price point that excludes most independent developers and limits use to research labs or companies with workstation budgets.
This doesn't mean the model is useless outside the two-Mac-Studio scenario. Meta has already released a hosted version via API, available through providers like Together AI, Groq, and OpenRouter. For the vast majority of developers, this will be the real path to adoption — the same strategy GitHub Copilot pursued when local IDEs couldn't scale.
The unlikely Meta-Apple alliance
The deal is as symbolic as it is technical. In 2021, Apple introduced iOS privacy changes that restricted user tracking — a direct blow to Meta's business model, which depends on targeted ads. The relationship reached rock bottom, with Zuckerberg saying at the time that Apple was "betraying" developers. Alexandr Wang, CEO of Meta AI, summed up the current state with one phrase: "I'm happy that mommy and daddy are playing nice again."
Behind the irony lies a clear financial motive. The race for closed frontier models is pushing inference costs to unsustainable levels. OpenAI, which reported an annualized revenue of $20 billion, faces investor pressure to reach its $50 billion target. Meta, whose stock has fallen 10% this year, needs to show that its investments in artificial intelligence deliver returns. The open-weight model allows Meta to distribute the inference cost to users — or, in the case of Muse Spark Open, to API providers that rent hardware on demand.
The partnership with Apple is another piece of the puzzle. The MLX framework was created to keep Apple's developer ecosystem locked into Apple hardware. By optimizing the weights of its biggest competitor for MLX, Meta gains native compatibility with the most powerful personal AI hardware available today; Apple, in turn, reinforces the argument that the Mac is the premier platform for personal AI.
The Chinese context and the geopolitics of open models
Meta positions Muse Spark Open as a direct response to Chinese advances in open-weight models. Startups like Moonshot AI (Kimi K3, 2.8 trillion parameters), Alibaba (Qwen3.8-Max, 2.4 trillion), and DeepSeek (V4-Flash) already offer systems that compete in quality with leading American models, but with permissive licenses that allow commercial modifications without restrictions.
Zuckerberg, in recent remarks, argued that foreign labs hold real advantages: looser restrictions on training data, fewer barriers to distillation (the technique of training smaller models using the output of larger ones), and less regulatory scrutiny on safety. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," he said.
This reflects a broader trend. While OpenAI and Anthropic keep their most advanced models as black boxes — accessible only via API, with prices scaling per token — Chinese companies have adopted open-weight as a market strategy. The result is an ecosystem where Chinese models dominate open-source segments (GitHub Copilot already accepts local model extensions) and competition for closed-API markets becomes increasingly a game of costs.
What to expect next
Muse Spark Open is a milestone, but not the end of the line. Meta has confirmed that benchmarks have not been independently verified — meaning the claim of "most capable U.S. model" still lacks external validation. The Muse Model License carries limited commercial terms; developers need to read the fine print before shipping products built on the weights.
The central question is: how long until an 816-billion-parameter open-weight model runs on a single consumer device? The answer probably won't come from Meta or Apple, but from a specialized chip — perhaps Groq's, perhaps another inference startup's — that finds a way to compress a terabyte of unified memory into a form factor that fits an affordable datacenter.
Until then, the Meta-Apple partnership on Mac Studios serves as proof of concept and a reminder that, in artificial intelligence, hardware architecture is as important as model architecture. And for now, whoever controls memory controls the future.
Sources: The Verge, CNBC, Technology.org
✓ Independent sources cross-checked and verified before publishing
