Meta releases Muse Spark Open, with more control and a high hardware bar

Meta released Muse Spark Open on October 15, making a version of its coding and agent model available beyond a hosted service. Reuters independently reported the release. For developers, the consequential change is the option to run and adapt the model themselves, subject to access requirements and its license. Reuters

There is a concrete artifact behind the announcement: Meta's Hugging Face repository lists model-weight shards and configuration files. But the repository is gated. Visitors can inspect the file listing without being able to download the weights. Bright checked that listing, not a downloaded or running model. Repository files

What it takes to run

Meta's model card describes 816 billion total parameters, with 42 billion active, and a one-million-token context window. Its stated GPU-memory requirements are at least 700 GB for the NVFP4 version and 2 TB for BF16. Model card

That puts self-hosting well beyond an ordinary laptop setup. The smaller active-parameter figure should not be mistaken for the memory needed to hold the model.

Meta also describes an MLX configuration using two M5 Ultra Mac Studios, and reports on-device coding and video-editing demonstrations. These are company-reported demonstrations, not Bright tests. The announcement lists hosted inference partners too, offering a separate way to try the model without operating the hardware. Meta announcement

For a team evaluating deployment, the useful question is how much reliable work a specific configuration completes for its total cost. Fitting the weights into memory does not establish acceptable speed, task reliability or savings.

What “open” permits

The Muse Model License allows modification and many commercial uses, subject to conditions. It requires attribution, including “Built with Muse Spark Open,” and compliance with its usage policy. Model-as-a-service businesses whose aggregate gross revenue, including affiliates, exceeds US$20 million in any consecutive twelve months must notify Meta and obtain a separate agreement before offering this model or its derivatives. That condition excludes internal use, research and non-model-as-a-service activities, including qualifying product features. License

Bright describes this as open-weight, not open-source. Access to weights alone does not establish open-source status; restrictions on fields of use conflict with the Open Source Initiative's definition. Open Source Definition

Why it matters

Bright's assessment is that the immediate opportunity is deployment choice. A team with sufficient resources can evaluate a self-hosted system, choose how it integrates with its tools and reduce reliance on a single hosted endpoint. Keeping inference local may help control data flows, but privacy still depends on the surrounding software, logging and permissions.

The limits matter just as much. Bright has not independently tested performance, reproduced the desktop demonstrations or established a cost advantage. Meta's capability claims should be treated as claims pending independent evaluation.

The next useful evidence is practical: reproducible setup instructions, measured speed on specified hardware, task completion rates and operating costs. Those results will show how far this release turns access to weights into useful access to AI.