Meta has released the weights for what it describes as the most capable open-weight artificial intelligence model built in the United States — and, in a move almost nobody predicted, it engineered the model to run on a Mac. Announced on Thursday, 15 October 2026, Muse Spark Open packs 816 billion parameters and a 1-million-token context window into a natively multimodal system built for complex coding work and agentic tasks. Its headline distinction, though, is hardware: Meta designed it to run locally on Apple’s computers, and Apple itself helped make that happen.
The announcement lands just two days after Meta unveiled Muse Wonder, its closed frontier model — a launch this site has already covered. Thursday’s news is a different proposition entirely. Where Wonder is a proprietary system accessed through Meta’s own channels, Spark Open ships its weights to the world: developers can download the model, run it themselves and build on top of it. And the headline partnership — Meta and Apple, two companies whose relationship has been defined by bruising fights over privacy and advertising — makes this one of the more startling alliances in recent technology history.
The headline numbers
On paper, Muse Spark Open is a heavyweight. The model uses a mixture-of-experts architecture with 816 billion total parameters, of which 42 billion are active for any given token — a design that keeps inference costs manageable relative to the model’s size. It accepts text, images, video and audio as inputs, making it natively multimodal rather than a text model with vision bolted on. Its context window stretches to 1 million tokens, enough to hold entire codebases or very long documents in a single prompt.
The compressed NVFP4 version of the weights weighs in at roughly 493 gigabytes — which already tells you this is not a model for a laptop. Meta says its own benchmarks show strong performance on agentic and tool-use tasks, the multi-step workflows where an AI model plans, calls tools and iterates. On pure coding benchmarks, early summaries suggest it trails some leading Chinese open models, a caveat worth noting for developers weighing their options.
Meta is distributing the weights broadly. Downloads are available through Amazon SageMaker AI, Google Cloud’s Model Garden, Hugging Face and Microsoft’s Foundry Models, while hosted API access is offered through providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. For most developers, the hosted route will be the practical one — the local Mac deployment is aimed at a much narrower, better-funded audience.
A rare alliance: Meta and Apple
The genuine surprise of the launch is the Apple partnership. Meta chief executive Mark Zuckerberg announced on X that Muse Spark Open’s weights were releasing that day and that the model runs well on new Mac Studios thanks to work with Apple’s John Ternus and his team. Apple’s chief executive replied publicly, congratulating Meta and praising how efficiently a model of that scale runs from a desk, adding that Mac Studio keeps pushing the boundaries of on-device AI.
Apple’s contribution is technical and substantial. The company optimised the model for its MLX machine-learning framework so that inference can be distributed across two to four linked Mac Studios, which pool their unified memory and communicate over Thunderbolt 5 and RDMA interconnects. In Apple’s own testing, four M5 Ultra Mac Studios generated 36.73 tokens per second using an 8,192-token context and producing 128 output tokens — a respectable clip for a model this size running entirely on local hardware, with no cloud servers involved.
Why would these two rivals cooperate? The incentives line up neatly. Apple gets a flagship demonstration that its silicon and MLX framework can run frontier-scale AI locally — a powerful selling point for the Mac Studio line. Meta gets proof that its open-weight strategy can deliver where closed models cannot: nobody is running a closed frontier model on a desk. Neither company would say whether the collaboration extends beyond this release.
The hardware catch
Running Muse Spark Open locally demands an enormous amount of memory: at least 1 terabyte of combined RAM. No single Mac can hold the model, so the setup calls for either two Mac Studios with 512 gigabytes of unified memory each, or four machines with 256 gigabytes apiece.
There is a timing wrinkle. Apple supports up to 512 gigabytes of unified memory per Mac Studio, but that configuration was only scheduled to arrive in late October 2026 — meaning the four-machine, 256-gigabyte setup is the configuration that works today. Either way, the price tag is professional-grade. The M5 Ultra Mac Studio starts at roughly Rs 6,29,900 in India, so four base units alone cost well over Rs 25 lakh before memory upgrades are factored in.
This is clearly not a consumer product. Meta is aiming at developers, research labs and companies that handle sensitive code or data and would rather keep it on premises than send it to a cloud API. For that audience, the ability to run a frontier-class open model behind their own firewall — auditable, customisable and free of per-token charges — is the entire point.
What Meta says it can actually do
Meta paired the model with Muse Code, its AI coding terminal, and put it to work on real jobs. In one demonstration, the model tackled a Mastodon client codebase, delivering bug fixes, new features and a full build across an agentic workflow. In another, Meta deployed it inside DaVinci Resolve, the professional video-editing suite, where it reviewed footage and performed cropping — including cutting vertical clips from longer material, all on local hardware.
The emphasis is on agentic labour: tasks that require the model to take many steps, use tools and check its own work, rather than simply answering a question. That positioning reflects where Meta sees the open-weight battleground heading — not chatbots, but autonomous coding and media workflows that developers can run and inspect themselves.
