Meta has released Muse Spark Open, an open-weight AI model it calls the “most capable” open-weight model in the United States. According to TechCrunch AI, Meta published evaluation results on Thursday showing that the model beats other popular open-weight models. It’s also built to run locally on Apple’s Macs, which is unusual for a model this size.
📌 Key points
- What it is: an open-weight model with 816 billion parameters and a 1 million token context window
- What it’s for: complex coding and agentic work, meaning tasks where the AI takes multiple steps on its own. It’s natively multimodal.
- Where it runs: on Macs, but only with at least 1 terabyte of RAM, per Engadget. Meta says a Mac Studio with the M5 Ultra chip is the best fit.
- How to get it: available now through several API platforms
- Why it matters: an American lab is competing hard in open weights again, and serious local AI hardware just got a flagship model to run
🧠 What Meta actually shipped
The headline numbers are big. With 816 billion parameters, Muse Spark Open sits at the heavy end of open-weight models. The 1 million token context window means it can take in an entire codebase or a long stack of documents in one pass.
Meta also showed it working on real tasks. It paired the model with Muse Code, its AI coding terminal, and pointed it at a Mastodon client. The company says the model handled “bug fixes, new features, [and] full build.” In another demo, it ran DaVinci Resolve, the professional video editing app, and reviewed and cropped footage.
What stands out is the second demo. Coding agents are everywhere now. A model that can drive a desktop creative app like Resolve shows Meta wants it doing general computer work, not just writing code.
💻 The Mac angle, and the catch
Designing a frontier-scale model for Macs makes sense. Apple Silicon uses unified memory, so the CPU and GPU share one large pool of RAM. That’s why high-end Mac Studios have become a favorite for people running big models at home or in small offices.
The catch is the price of entry. A 1TB RAM requirement rules out nearly every machine people actually own. For now, “runs on a Mac” really means “runs on a top-spec Mac Studio.” Most developers will reach Muse Spark Open through an API, at least to start.
The direction still matters. The ceiling for local AI keeps going up. Model sizes that once needed a rack of data center GPUs now fit on one desktop machine, provided you’re willing to pay for it.
🌍 Why this matters for the industry
Meta’s “most capable in the United States” wording is deliberate. Over the past couple of years, Chinese labs have released many of the strongest open-weight models, while most leading US labs kept their best work closed. Meta built its open-weight reputation with the Llama family, and this release puts it back in that race.
Open weights matter for practical reasons:
- Cost: running the model on your own hardware or through competing API hosts can be much cheaper than closed models
- Control: you can fine-tune it, inspect it, and deploy it on your own terms
- Privacy: local inference keeps sensitive code and data on machines you control
- No lock-in: you’re not tied to one vendor’s pricing or policy changes
🔍 What to watch next
A few things will decide whether Muse Spark Open really takes hold:
- Independent benchmarks. These are Meta’s own evaluation results. Third-party testing on coding and agent tasks will tell us much more.
- License terms. “Open-weight” doesn’t automatically mean permissive. Check the commercial usage rules before you build on it.
- Smaller variants or quantized versions. Quantization compresses a model so it needs less memory, usually with some loss in quality. Builds that run on 256GB or 512GB machines would bring in far more users.
- API pricing. Since most people will use it through hosted platforms, per-token costs will shape how widely it gets adopted.
My take for teams: if you’re spending a lot on closed models for coding agents, put Muse Spark Open on your evaluation list now. Test it through an API before you even think about buying hardware.
TechCrunch AI has more details on the launch and Meta’s evaluation results in its original report.