Is Muse Spark Open Source? What Meta's License Allows

No. Muse Spark Open is an open-weights model that Meta released under its own Muse Model License, which is not an open-source license approved by the Open Source Initiative (OSI). At Layer3Labs, we build and run AI systems inside other people's businesses, and for this release the license is a smaller hurdle than the hardware. The license lets you download, fine-tune and sell the model free of charge. The license sets two conditions: a "Built with Muse Spark Open" notice, and a separate agreement with Meta for model-as-a-service businesses above US$20 million in revenue.

Meta released the weights for Muse Spark Open on October 15, 2026. At the same time, Meta kept its frontier models, Muse Spark 1.3 and Muse Wonder, restricted to proprietary Application Programming Interface (API) access.

Is Muse Spark Open Source? Open Weights With a License

Muse Spark Open is an open-weights model governed by the custom Muse Model License, making it fundamentally different from traditional open-source software. Meta released the weights on October 15, 2026, after development by Meta Superintelligence Labs. On research.meta.ai, Meta described the release as the open weights version of Muse Spark and called it the most capable US open-weight model to date. Independent testing has not verified that claim. Muse Spark 1.3 and Muse Wonder, by contrast, remain closed-weight and API-only.

A post on the PyTorch blog explains that Muse Spark Open builds on Meta's earlier August release of Muse Spark 1.2. The PyTorch post says it shares the same base model as Muse Spark 1.3. Meta gave the open version extra safety post-training so it is more robust against misuse as a standalone model without system guardrails. While several press reports refer to the model as Muse Spark 1.2, Meta's own pages use the name Muse Spark Open.

The Muse Model License is not an OSI-approved open-source license, and Meta itself describes it as having "limited commercial terms". The industry uses the term open weights to describe downloadable models with custom legal terms. To understand how these models function across enterprise stacks, read our guide on what open-weights models are. You can review the exact legal clauses within the LICENSE.md file located in the official Hugging Face repository.

• Open weights released on October 15, 2026 by Meta Superintelligence Labs
• Shares the same base model as Muse Spark 1.3, with extra safety post-training for standalone use
• Governed by the Muse Model License rather than an OSI-approved open-source license
• Meta claims leading US capability, though independent benchmarks have not verified this

What the Muse Model License Lets You Do

The Muse Model License grants royalty-free rights to use, modify, fine-tune, and distribute Muse Spark Open, but it enforces strict attribution rules and a revenue threshold for Model-as-a-Service businesses. Under the basic grant, you can copy, modify, create derivatives, distill, merge, publish, distribute, sublicense, and sell copies of the software. Meta provides the model software on an as-is basis without express or implied warranties, and Meta claims no ownership over outputs your systems generate.

Attribution is mandatory across all distributions. You must retain the copyright notice, license text, and Usage Policy in any copy or derivative model you produce. Furthermore, you must prominently display the phrase "Built with Muse Spark Open" within your product user interface, documentation, website, or settings menu. Derivative models include any model trained, fine-tuned, or distilled using the weights or model outputs, and they carry the same notice and Usage Policy duties.

The revenue restriction applies exclusively to Model-as-a-Service (MaaS) providers offering third parties direct API access to model outputs. If your business and its affiliates generate more than US$20 million in aggregate gross revenue over any consecutive 12-month period, you must notify Meta and execute a separate written agreement. Internal business operations, academic research, and standard software products that incorporate the model do not trigger this clause. The Muse Model License sets no monthly-active-user threshold.

Licensees must also comply with Meta's Acceptable Use Policy, dated October 15, 2026. The policy states the software is not intended for users under 18 years of age. Prohibited uses include criminal conduct, harassment, malware development, circumventing safety measures, weapons and military use, attacks on critical infrastructure, fraud, disinformation campaigns, unauthorized impersonation, and presenting automated model outputs as human-written text.

What Meta Released in Muse Spark Open

Meta structured Muse Spark Open as an 816-billion-parameter sparse mixture-of-experts model built for coding and agentic work. The architecture activates approximately 42 billion parameters per token across 256 individual experts. All 816 billion parameters still have to sit in memory, so even the smallest community build needs about 300 GB of combined system and graphics memory.

The context window is 1 million tokens, and Meta's vLLM example sets a maximum model length of 1,048,576 tokens. For input data, Muse Spark Open natively processes text, image, video, and audio inputs. Outputs consist entirely of text. Meta built Muse Spark Open for agentic work: long-horizon tasks, tool calling, subagents, web browsing and computer use. One coding agent can read a screenshot, recording or long codebase and act on it in the same run.

Meta's official announcement links to an evaluation methodology report, but the publication provides no specific benchmark scores. Additionally, Meta noted that the full safety evaluation report remains unpublished and will become available soon. Because vendor benchmark numbers are absent, engineering teams must evaluate the model directly on internal test suites to verify its accuracy and reliability.

• Total parameter count of 816 billion with 42 billion active parameters per token
• Sparse mixture-of-experts architecture routing tokens across 256 specialized experts
• Native multimodal input processing for text, images, video, and audio streams
• Native context window of 1 million tokens with text-only generation output

Which Muse Models Stay Closed

Meta maintains a strict boundary between its open-weights distribution and its most capable proprietary models. While developers can download Muse Spark Open, Muse Spark 1.3 and Muse Wonder remain closed-weight models accessible only through Meta's commercial Application Programming Interface (API). Press coverage reports that Meta Chief AI Officer Alexandr Wang has said Meta will generally not open-source its most powerful models.

Meta's earlier open release, Muse Glimmer, is a small dense model.

Muse Spark Open and the smaller Muse Glimmer are the two Muse models you can run on your own hardware. To see how other US labs handle open and closed releases, examine our guide to American open-weight AI models and our overview of the broader Meta Muse architecture.

• Muse Spark 1.3 remains closed-weight and restricted to proprietary API endpoints
• Muse Wonder remains a closed-weight, API-only model with no weights available
• Muse Glimmer is Meta's earlier small dense open model
• Press coverage reports Alexandr Wang said Meta will generally not open-source its most powerful models

Where to Get It: Download or Hosted API

Developers can obtain the official Muse Spark Open model files directly from the Hugging Face repository under the namespace meta-models/Muse-Spark-Open-816B-A42B. Meta published checkpoints in both 16-bit Brain Floating Point (BF16) and 4-bit NVIDIA Floating Point (NVFP4) formats. Cloud marketplaces including Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models also list the official weights. Community groups have published quantized GGUF format builds under unsloth/Muse-Spark-Open-816B-A42B-GGUF and ggml-org/Muse-Spark-Open-816B-A42B-GGUF.

If your organization prefers managed cloud hosting over physical infrastructure, Meta identified several inference providers hosting Muse Spark Open endpoints. Supported partners include Baseten, Fireworks AI, Groq under a limited preview, Modal, Nebius, OpenRouter, and Together AI. Meta published no per-token pricing for these services, so cost comparisons depend entirely on the billing schedules each provider establishes.

Meta verified local inference on multiple serving frameworks, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth across NVIDIA, AMD, Intel, and Apple silicon. Meta provides no official Ollama package. Self-hosting needs a lot of memory, because the unquantized BF16 checkpoint spans roughly 1.6 TB, while the NVFP4 format requires approximately 530 GB. For complete server configurations, memory requirements, and cluster guidance, read our guide on how to run Muse Spark Open locally.

• Download official BF16 and NVFP4 checkpoints directly from Hugging Face
• Find the weights in Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models
• Access serverless endpoints through providers like Baseten, Modal, and Together AI
• Run self-hosted inference using vLLM, SGLang, TensorRT-LLM, or Apple MLX

Who Should Adopt Muse Spark Open Now

Adopting Muse Spark Open makes sense for organizations with strict compliance mandates, data residency rules, or dedicated machine learning engineering teams. In our view, if your workflows must keep sensitive customer data on your own servers, self-hosting these weights keeps prompts and outputs off third-party APIs.

Teams building coding agents get the most from it, because Meta built Muse Spark Open for long-horizon coding, tool calling and subagents, and demoed Muse Code running on it on-device. Organizations can adapt the weights using TorchTitan for Supervised Fine-Tuning (SFT), Low-Rank Adaptation (LoRA), and Reinforcement Learning (RL). NVIDIA NeMo AutoModel handles full SFT and LoRA, while Hugging Face SFTTrainer provides standard adaptation pipelines. Managed fine-tuning is also available through Fireworks AI and Together AI. Review our guide on fine-tuning open-weights models to determine whether fine-tuning or retrieval fits your specific use case.

Who should wait: Teams that require independently audited performance benchmarks, published safety verifications, or a warranty (Meta provides the software "as is") should hold off on production rollouts. Meta has not yet released the full safety evaluation report. If your compliance policies require formal security certifications, wait until third-party labs publish safety audits. For security considerations, read our analysis on whether open-weights models are safe. To compare Muse Spark Open against alternatives, see our index of the best open-weights AI models or the Kimi K3 open-weights guide.

Who this is not for: This release is not suitable for small teams without dedicated machine learning infrastructure or multi-GPU hosting. If your team lacks cluster administration experience, query a managed cloud endpoint instead of attempting to host an 816-billion-parameter cluster locally. What would change our answer: if independent benchmarks show clear operational advantages over hosted frontier models, we would recommend broader deployment across mid-sized business automation stacks. We would also widen that advice if community quantizations reduce latency on consumer-grade hardware without degrading reasoning.

• Adopt now if data residency, on-premise control, or deep fine-tuning are mandatory
• Adopt now for agentic coding tools, tool calling, and long-context multimodal evaluation
• Wait if your organization requires published third-party benchmarks or safety audit reports
• Avoid self-hosting if you lack multi-GPU infrastructure or specialized cluster engineers

What you need to run Muse Spark Open yourself

Muse Spark Open is a frontier-scale Mixture-of-Experts model, so "running it yourself" is a real infrastructure decision — not something a single laptop or gaming GPU can do. Match the path below to how seriously you need to self-host. For most teams the API or rented GPUs are the right answer; buying hardware only pays off at steady, high volume or when your data can never leave your walls.

Once Muse Spark Open is running, the fastest way to put it to work day to day is inside Cursor — point it at the model through OpenRouter as a custom model. And if you would rather run a model on one affordable box, see Best mini PCs for local AI and Local AI hardware calculator.

NVIDIA RTX 6000 Ada (48GB)

Power users and small clusters that want cards they own

View on Amazon →