Meta’s Muse Spark Open Is an Open-Weight Model With Strings Attached

Meta is back in the open-weight race, and it wants credit for bringing a serious model with it. Muse Spark Open ships with 816 billion parameters, 42 billion active per request, a one-million-token context window, and a design aimed at coding and AI agents. Meta says it’s also the first model in its class that can run on a desktop, if that desktop is actually two linked Mac Studios with Apple’s M5 Ultra chip.

The model is natively multimodal, so it can work with images and video as well as text. Meta showed it fixing bugs in an iOS app by itself and going through footage frame by frame inside DaVinci Resolve. The weights are on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and the model is also available through API providers including Together AI, Fireworks, Groq and OpenRouter.

But the practical cost is steep. Unite.AI says the model needs at least 700 gigabytes of graphics memory depending on the format, and the two Mac Studios with 512 gigabytes each are priced at roughly what a compact car costs. So yes, it’s open-weight. No, it’s not casual.

Meta is also making a strategic turn here. With Muse Spark, the company had already introduced its first model without open weights in the spring, breaking from the old Llama playbook. Now Muse Wonder puts Meta in the closed frontier-model race too, while Muse Spark Open is meant to restore some of the open-weight credibility.

On performance, the picture is mixed and still mostly self-reported. Meta says Muse Spark Open beats American rivals such as Nvidia’s Nemotron 3 Ultra and Thinking Machines’ Inkling by more than 20 points on some benchmarks, but it still trails the strongest Chinese open-weight models on several tests, including the Artificial Analysis Intelligence Index, Terminal-Bench 2.1, GPQA Diamond and SWE-Atlas-QnA. In Meta’s own data, the model leads on agent and tool tasks like JobBench and MCP-Atlas, and it runs neck and neck with Mistral Large 4 on the Artificial Analysis score. None of this has been independently confirmed yet, and Muse Spark Open is not listed in Artificial Analysis’s ranking of open-weight models.

The license is where the “open” label gets fuzzy. Meta’s Muse Model License allows free use, copying, distribution, fine-tuning, derivative models and distillation, and companies can use it internally or inside their own products, including commercially. But if someone wants to run it as an AI service with more than $20 million in revenue over 12 months, Meta wants advance notice and a separate agreement. Products using the model also have to show “Built with Muse Spark Open,” and the acceptable-use policy bans weapons development, malware, critical infrastructure, fraud, disinformation and impersonation. For most developers that’s manageable. For the biggest cloud and AI providers, it means Meta still holds the keys.