Meta releases Muse Spark Open, an 816B model with restricted commercial terms

Meta released Muse Spark Open on October 15, 2026, introducing a massive 816 billion parameter model designed for complex agentic tasks. The company positions this release as the most capable open-weight model from the United States, aiming to match or exceed the performance of leading Chinese alternatives while maintaining a distinct licensing structure.

What happened

The new model features 42 billion active parameters and supports a context window of one million tokens. Meta made the weights available for download through major platforms including Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft’s Foundry Model Catalog. Developers can also access the model via APIs from providers such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

While the weights are publicly accessible, the accompanying Muse Model License imposes specific commercial restrictions. Companies generating more than $20 million in gross revenue over any consecutive twelve-month period must notify Meta and enter into a separate written agreement before offering the model or its derivatives as a service. This threshold effectively limits unrestricted commercial deployment to smaller entities and internal enterprise use.

The license also mandates that any product or service built using Muse Spark Open must prominently display the attribution "Built with Muse Spark Open" in its user interface. These terms mirror restrictions seen in recent licenses from Chinese competitors like Moonshot for Kimi K3. The policy explicitly bars military and critical infrastructure deployments, a constraint Meta previously introduced for its Llama models in 2024, but it does not include the broader usage prohibitions found in some competitor policies.

How it works

Muse Spark Open utilizes a sparse architecture where only 42 billion of its 816 billion parameters are active during any given inference step. This design allows the model to maintain a vast knowledge base and reasoning capability while reducing the computational cost per token compared to dense models of similar total size. The one million token window enables the model to process extensive codebases, long documents, or multi-step agent workflows without losing context.

Meta optimized the model for coding and agentic work, focusing on long-horizon tasks and multi-step reasoning. To demonstrate local viability, Meta collaborated with Apple to run the model on two Mac Studio units equipped with M5 Ultra chips and 512 GB of memory each, linked together using the MLX framework. This setup allowed Meta to power Muse Code and ship updates to the Ice Cubes Mastodon client, though such hardware remains prohibitively expensive for most individual developers.

Key details

• Model size: 816 billion total parameters with 42 billion active parameters.
• Context window: Supports up to one million tokens.
• Revenue threshold: Companies with over $20 million in annual gross revenue must sign a separate agreement for service offerings.
• Attribution requirement: Products must display "Built with Muse Spark Open" in the user interface.
• Availability: Weights are on Hugging Face, SageMaker, Google Cloud, and Microsoft Foundry; APIs are available via eight major providers.
• Hardware demo: Ran locally on two Mac Studios with M5 Ultra chips and 512 GB RAM each.

Why it matters

For software engineers and technical founders, Muse Spark Open represents a high-performance option for agentic workflows that rival Chinese models like Kimi K3, GLM 5.3, and Qwen3.8. In benchmarks focused on tool use and agentic tasks, such as JobBench and MCP-Atlas, Muse Spark Open leads these competitors. However, it still trails in pure coding benchmarks, scoring 76.4 on Terminal-Bench 2.1 compared to Kimi K3’s 85.0. This suggests the model is best suited for complex, multi-step reasoning tasks rather than raw code generation alone.

The licensing structure introduces a clear compliance boundary for growing startups. While internal use and integration into proprietary products remain free for all, crossing the $20 million revenue threshold triggers a negotiation requirement for service-based offerings. This creates a predictable path for scaling but adds legal overhead for successful companies. The mandatory attribution clause also impacts product design, requiring visible branding that may not align with all white-label strategies.

What you can do

• Evaluate Muse Spark Open for agentic tasks and long-context reasoning where it outperforms US rivals like Thinking Machines’ Inkling.
• Review your company’s gross revenue against the $20 million threshold to determine if a separate agreement with Meta is required for commercial service offerings.
• Implement the required "Built with Muse Spark Open" attribution in your user interface if you deploy the model or its derivatives.
• Test the model via API providers like Fireworks AI or Groq before committing to local infrastructure costs.
• Consider the coding performance gap if your primary use case is standalone code generation, as it trails behind Kimi K3 and GLM 5.3 on relevant benchmarks.
• Monitor hardware requirements if planning local deployment, noting that optimal performance currently demands high-end setups like dual Mac Studios with 512 GB RAM.