Meta Claims Crown for Open AI With 816-Billion-Parameter Model

Meta launched an 816-billion-parameter model on Thursday and called it the "most capable" open-weight AI model available in the United States. Muse Spark Open, as the company named it, also carries an unusual hardware target: it is designed to run locally on Apple's Macs. That combination — record-setting size plus desktop deployment — puts Meta in direct competition with the open-weight field it now claims to lead.

The claim rests on evaluation results Meta itself released. No independent benchmarks appear in the announcement. Readers should treat the "most capable" designation as a vendor assertion until third parties run the model.

What does the model actually do?

Muse Spark Open is natively multimodal, meaning it handles more than one input type, and Meta built it for two demanding workloads: complex coding and agentic labor — AI that acts on tasks with limited supervision. The 1 million token context window lets it process very long documents or codebases in a single pass.

Meta says it tested the model inside Muse Code, its own AI coding terminal, and deployed it across agentic workflows. In one disclosed example, the model handled coding fixes for a Mastodon client that included "bug fixes, new features, [and] full build."

The company also reported running DaVinci Resolve, the high-end video editing software, with the model capable of reviewing and cropping footage. Both examples come from Meta's own testing. Whether the performance holds up in outside production environments remains unverified.

Who can actually run it?

Here is the catch. Local operation requires at least 1 terabyte of RAM, per Engadget. For perspective, a well-equipped consumer laptop ships with 16 gigabytes — roughly one-sixty-fourth of that. Even high-end workstation configurations rarely exceed 128 gigabytes.

Meta recommends the Mac Studio equipped with the M5 Ultra chip for best results. That narrows the practical audience to developers and studios with server-grade budgets. For everyone else, cloud deployment is the realistic path.

The model is available now through a number of API platforms, Meta said. The company did not disclose pricing tiers in the announcement.

Why open weights matter commercially?

Open-weight models differ from closed, proprietary systems in one decisive way: anyone can download and inspect the model's parameters. That structure typically delivers three advantages Meta is leaning on:

• Lower cost at scale, since users avoid per-token licensing fees
• Transparency, because the weights can be audited
• Flexibility to modify and deploy the model on private infrastructure

Closed models from rivals keep their weights secret and sell access. Meta's strategy inverts that, betting openness will pull developers into its ecosystem.

The bet is not cost-free. Open weights mean anyone — including competitors — can build on Meta's work. The company appears willing to accept that trade to set industry standards on its own terms.

What to watch next?

Two questions will decide whether Thursday's launch matters beyond the press release. First, independent benchmarks must confirm Meta's performance claims against other open-weight models. Second, the 1-terabyte RAM requirement either falls as optimization improves, or the model stays a niche tool for Mac Studio owners.

Until both resolve, Muse Spark Open is a statement of intent: Meta wants to define the ceiling of what open AI can do, and it is willing to give the weights away to do it.