Meta wants its open model taken seriously. On October 15, 2026, the company launched Muse Spark Open, an open-weight AI model aimed at competing with products from leading Chinese developers and running on Apple hardware.
Meta says Muse Spark Open rivals, and in some cases outperforms, similar Chinese models. Meta’s chief of AI, Alexandr Wang, called it “the most capable US open-weight model to date,” a confident claim in a market where every model launch arrives wearing a crown.
The timing adds pressure to Meta’s broader model strategy. Meta announced its newest closed model just days before October 15, 2026, then introduced Muse Spark Open as an open-weight alternative with transparency and flexibility built into its appeal.
A model built for long tasks
Muse Spark Open uses 816 billion parameters and offers a 1 million token context window. It is natively multimodal, built for complex coding tasks, and designed to handle agentic labor rather than only answer isolated prompts.
Meta says the model excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Those goals place the emphasis on work that unfolds across several stages, where the system must maintain context, make decisions, and complete a larger assignment instead of producing one polished paragraph and calling it a day.
Meta’s evaluation results claim that Muse Spark Open delivers superior performance to other popular open-weight models. The company also positions it as a tool for coding and agentic work, giving the model a role in workflows that involve software changes, testing, and other connected tasks.
Meta paired Muse Spark Open with Muse Code, its AI coding terminal, and deployed it in workflows that included coding fixes for a Mastodon client. Those fixes included “bug fixes, new features, [and] full build,” according to Meta — a useful demonstration of scope, assuming the model’s performance holds outside the announcement.
Mac hardware becomes part of the pitch
Muse Spark Open is optimized to run on Apple’s advanced M5 Ultra chips. Meta says the model will run best through a Mac Studio equipped with the M5 Ultra chip, turning Apple hardware into part of the model’s story rather than a footnote beneath it.
Users can run Muse Spark Open on two Mac Studios with M5 Ultra chips to create a supercomputer on their own desktop. The setup requires at least 1 terabyte of RAM, so this is not a casual experiment for an ordinary workstation — though “supercomputer on your desktop” is exactly the sort of phrase designed to make expensive hardware sound like a personal breakthrough.
The hardware requirement also shows the trade-off behind a model with 816 billion parameters and a 1 million token context window. Open-weight access can offer flexibility and transparency, but running a system of this size still depends on serious computing resources.
Meta also deployed Muse Spark Open in the service of running DaVinci Resolve, the high-end video editing software. The company says the model could review and crop footage, extending its use beyond software development into video workflows.
That combination of coding, software operation, and video handling supports Meta’s claim that Muse Spark Open is multimodal and suited to agentic work. It also gives the model a broader test than a benchmark score: the system must interact with tools and carry out tasks across different kinds of media.
Open-weight models can be more cost-efficient and offer transparency and flexibility. Muse Spark Open turns those advantages into a direct challenge to Chinese models, closed systems, and the assumption that serious AI work must happen behind a company’s private interface.
Meta’s claims still come from Meta, and the company’s evaluation results do not settle every question about real-world performance. But Muse Spark Open brings together a large model, long context, multimodal capabilities, coding tools, video editing, and Mac-based deployment in one package — which gives the open-weight market something more useful than another leaderboard trophy.
Based on
