Meta ships Muse Spark Open, its answer to China’s best open-weight models

Muse Spark Open is Meta’s new open-weight AI model, with 42 billion active parameters and a one-million-token context window. Meta presents it as the most capable U.S. open-weight model so far. Its focus is coding, agentic work, long tasks, software engineering, and multi-step reasoning.

Developers can access the model through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. They can also download its weights from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft’s Foundry Model Catalog.

Access is broad, but the license is not completely unrestricted. Companies with more than $20 million in gross revenue during any consecutive twelve-month period must notify Meta and sign a separate agreement before offering model-as-a-service products. Users must also display “Built with Muse Spark Open.”

Muse Spark Open contains 816 billion total parameters, the adjustable numerical values that store patterns learned during training. However, Meta identifies only 42 billion as active parameters. The distinction matters because the model’s full capacity is much larger than the portion involved in each individual response.

In practical terms, total parameters describe the model’s overall installed knowledge and capabilities, while active parameters describe how much of that capacity is engaged for a particular request. This design can reduce the computation needed per response compared with using all 816 billion parameters every time. The article does not name the specific routing method.

The model remains extremely large for local use. Meta says two Mac Studios with M5 Ultra chips and 512 GB each could be linked to run it with MLX. Most developers will therefore use hosted APIs or downloadable weights on suitable infrastructure instead.

Against leading Chinese open models, Muse Spark Open performs best on agentic work and tool use. Meta reports that it leads Nyaya K3, Cosmos, and Qwen3.8 on benchmarks including JobBench and MCP-Atlas. It roughly ties those models on SciCode, showing strength in scientific coding tasks.

Coding results are less favorable. On Terminal-Bench 2.1, Nyaya K3 scores 85.0 and Cosmos scores 83.9, compared with 76.4 for Spark Open. On DeepSWE 1.1, the spread is 68 to 60, with Spark Open behind. These results point to a specific weakness in software-engineering performance.

Meta’s model still compares strongly overall, especially in long-horizon and tool-using tasks. But the article says the coding gap resembles the one acknowledged when the closed Spark model launched in April, suggesting limited progress since then.

Muse Spark Open uses Meta’s Muse Model License, which Meta describes as having limited commercial terms. Internal use and building the model into a company’s own product are free. However, the license places extra requirements on companies with substantial revenue.

Any company earning more than $20 million in gross revenue during any consecutive twelve-month period must notify Meta and enter a separate written agreement before offering the software or derivatives as part of a model-as-a-service business. This applies when the model is provided as a service to customers.

Every company using the model must also prominently display “Built with Muse Spark Open” in the interface of each product or service using the model or a derivative. The policy bars military and critical-infrastructure deployments. The article says developers can abuse the model freely, unlike Paradox’s newer Paradox usage policy.

A model’s weights are the files containing the learned numerical settings used to generate outputs. Making those files available gives developers significant access. But “open weights” describes availability, not necessarily the freedoms granted by the license attached to them.

Muse Spark Open illustrates the difference. Developers can download its weights, use them internally, and build them into products. Yet companies above $20 million in gross revenue must notify Meta and sign an agreement before offering model-as-a-service products. Users must also show “Built with Muse Spark Open,” and military and critical-infrastructure deployments are barred.

Those conditions make the model less permissive than a model with unrestricted commercial and deployment rights. The article calls the license semi-open in effect and compares it with Moonshot’s license for Nyaya K3. The practical question is therefore not only whether weights are downloadable, but also what the license allows.

An 816-billion-parameter model is difficult to run locally, even though Meta says a particular two-computer setup can handle it. Developers who lack that hardware do not need to operate the model themselves. They can send requests to services that host the model on their infrastructure.

The article lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. These providers handle the computing burden, while developers integrate the model into applications through their services. The model’s weights are also available through several platforms for teams with their own suitable infrastructure.

This creates two practical paths: hosted access for convenience, or self-hosting through downloaded weights when resources permit. Meta’s example shows the local hardware challenge clearly: two new Mac Studios with M5 Ultra chips and 512 GB each were linked with MLX to run the model. Two such machines may cost about as much as a compact car.

Parameters are adjustable numerical values that a model changes during training to capture patterns in data. Weights are commonly used for those learned values and the files that store them. More parameters can provide more capacity for complex patterns, though size alone does not guarantee better results. The article identifies 816 billion total and 42 billion active parameters for Muse Spark Open.

A token window is the amount of text a model can process as context in one interaction. Muse Spark Open has a one-million-token window, so it can consider very large prompts or documents at once. That helps with long tasks, but it does not mean the model permanently remembers every conversation.

During generation, the model uses its weights and the supplied context to predict an output step by step. The article connects Spark Open with long-horizon tasks and reasoning, but it does not explain its training process or memory behavior in detail.