Muse Spark Open

Strengths

• Open weights of a large model (816 billion parameters, 42 billion active) with a 1M-token window. It understands text, images, audio and video.
• The license does not exclude the EU, and Meta Platforms Ireland is the licensor for Europe. When you run it yourself, data stays with you.
• Commercial use with no user cap. Only providers that sell the model as a service and earn over $20M a year need a separate agreement with Meta.
• Broad support from day one: SageMaker, Google Model Garden, Microsoft Foundry, vLLM, SGLang, llama.cpp and MLX.
• Meta's model card shows it is good at operating a computer and tools: 80.4 on OSWorld Verified and 89.2 on MCP-Atlas.

Weaknesses

• It needs data-centre hardware: even in the compressed NVFP4 form the weights are about 537 GB (one server with 4 NVIDIA Blackwell GPUs), and full precision needs about 2 TB of GPU memory.
• It is not an OSI-approved open source license: products must show “Built with Muse Spark Open”, a usage policy applies and use is limited to people 18 and over.
• Clearly weaker than Meta's hosted models: 39 on the Artificial Analysis index, against 48 for Muse Spark 1.3.
• Through providers it costs the same per token as the hosted Muse Spark 1.3 ($1.25 and $4.25), so it is not cheaper via API.

Best for

• running on your own servers or a private cloud
• fine-tuning a large open model
• agents working on sensitive data that must not leave the company

Not ideal for

• a single GPU or a laptop, Muse Glimmer 30B is meant for that
• teams that need an OSI-approved open source license

How to use it in Europe

You can download the weights from Hugging Face after filling in a form, including from the EU. The Muse Model License has no EU carve-out, and Meta Platforms Ireland is the licensor for European users. Without your own servers, providers such as Nebius, Together and Fireworks host it. Running it yourself keeps data inside your infrastructure, which is the safest option for privacy.

Price compared with similar models

Benchmark results

Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.

More models from this vendor

• Muse Wonder: Meta's strongest model for long-running agents and coding, still released as a preview.
• Muse Spark 1.3: Meta's balanced model with a 1-million-token window and a fair price. The Muse Spark line also powers the Meta AI assistant.
• Muse Glimmer 30B: Meta's small open model under Apache 2.0 that runs on a single graphics card, even offline.
• Llama 4 Maverick: The larger model in the newest Llama generation, from April 2025. Cheap through providers, but its license excludes people and companies in the EU.
• Llama 4 Scout: The smaller Llama 4 from April 2025, very cheap through providers. Its license grants no rights to people or companies in the EU.

Similar models from other vendors

• Claude Sonnet 5.5 (Anthropic): Claude's balanced model for everyday work and coding. It is faster than Opus 5.5 and costs half as much.
• Command A+ (Cohere): An open business model from Canada's Cohere that officially supports Slovak and Czech and runs on your own server.
• Gemini 3.8 Flash (Google): Google's fast, lower-cost model for everyday work, coding and document analysis. You can try it in the API for free.
• Mistral Medium 3.5 (Mistral AI): Mistral's mid-size open-weight model and the default in the Vibe app. Good for everyday work and coding with data kept in the EU.
• Nemotron 3.5 Super VL (NVIDIA): NVIDIA's open model for business AI agents that understands text, images and video and is very fast and cheap to run.
• GPT-6.1 Sol (OpenAI): OpenAI's balanced model, close to GPT-6 Astra in quality at a fifth of the price. A solid pick for business apps.

• All AI prices in one place

Sources

• Meta docs: Muse Spark Open ↗
• Hugging Face: Muse Spark Open model card ↗
• Muse Model License ↗
• OpenRouter: Muse Spark Open providers ↗
• DeepLearning.AI The Batch: Meta releases Wonder, opens Spark ↗