Muse Spark Open: Meta Returns to Open-Weight Models

Meta is back in the open-weight business. After the Llama series, the company has released Muse Spark Open, a large model whose weights anyone can download. Meta calls it the most capable US open-weight model available. It is built for coding and agentic tasks, meaning work where the model plans and carries out multi-step jobs rather than answering a single prompt.

The release puts Meta in direct competition with Qwen, GLM and Kimi, the model families that have shaped the open-weight field recently. It also shows where Meta draws the line between what it gives away and what it keeps behind a paywall.

What Meta has released

The headline numbers:

• Size: 816 billion parameters in total, with 42 billion active at any given time.
• Context window: up to one million tokens, which is enough to hold very long documents or large codebases in a single session.
• Modalities: text, image, video and audio input.
• Focus: coding and agentic workloads.

The gap between total and active parameters matters. A model this large would normally be expensive to run. Because only around 5 percent of the parameters are used at once, the compute cost per request is much lower than the total size suggests.

How it compares

Meta points to several benchmarks: JobBench, MCP-Atlas, HLE-Diamond and ZeroBench. On these, the company says Muse Spark Open scores well ahead of Qwen 3.8, GLM 5.3 and Kimi K3 in several categories.

Two caveats apply. First, these are Meta's own results, and "several categories" is not the same as across the board. Second, Meta openly says the model falls well short of its paid Muse Wonder models, specifically the Muse Wonder Full tier. The open model is meant to lead its own class, not replace Meta's commercial top end.

Running it locally

Meta says the model is optimized for Apple's M5 Ultra chip. According to the company, it runs locally on two Mac Studios, each with 512 GB of unified memory. That is still serious hardware, but it is a desk-sized setup, not a data center rack.

For developers and companies that do not want to send code or internal data to an outside API, this is the most practical part of the announcement. A model of this size running entirely on hardware a team controls was hard to imagine not long ago.

The weights are available on Hugging Face, the platform most developers use to share and download AI models. Meta has also published a research blog with further technical detail.

The license terms

The model is free to use, with conditions. Two stand out:

1. Attribution: anything built on the model must carry the label "Built with Muse Spark Open."
2. Revenue threshold: providers earning more than $20 million a year who want to offer the model as an API service need a separate agreement with Meta.

The second rule targets cloud providers and large hosting companies. Small teams and most businesses using it internally are not affected. Companies that want to resell access to it at scale must negotiate with Meta.

Our Take

The release tells us where Meta wants to sit in the market. It keeps its strongest Muse Wonder models closed and paid, and uses an open model to win developers' attention. The labelling requirement and the API revenue threshold suggest Meta wants visibility and some control over how others profit from its work, even while it gives the weights away.

The more interesting angle is the "US open-weight" framing. Qwen, GLM and Kimi have set the pace for downloadable models, and Meta is not the only American company trying to compete. We recently covered Reflection Beam, which targets the same rivals, and Mistral has been pushing large open models from Europe with Mistral Large 4. This suggests origin is becoming a selling point in its own right, especially for organisations that care about where their models come from.

For readers who build with AI, the practical question is whether the benchmark claims hold up outside Meta's own testing. Independent evaluations over the coming weeks will show whether Muse Spark Open actually leads its class in coding and agent work, or only in the categories Meta chose to highlight. It is also worth watching how many teams take up the local setup on Mac Studio hardware, and whether the $20 million API rule keeps the large cloud providers away or pushes them into deals with Meta. Finally, the response from Qwen, GLM and Kimi will matter. Releases in this field tend to come quickly, and any lead Meta has now may not last long.