Meta releases Muse Spark Open as a desktop-ready open-weight model

Meta announced Muse Spark Open on October 15, 2026, and released its model weights immediately. The natively multimodal model has 816 billion total parameters, of which 42 billion are active for each token, and a context window of one million tokens. Meta positions it for coding, long-running agent tasks and multi-step reasoning. The release is a material change from the hosted Muse Spark 1.3 update covered earlier: developers can now download and adapt the weights instead of using only Meta’s managed surfaces.

The unusual hardware story is part of the announcement. Meta says Muse Spark Open is the first model in its class that can run on a desktop made from two Apple Mac Studios with M5 Ultra chips, using the MLX framework and their combined unified memory. Meta demonstrated the setup on an open-source Mastodon client, where the model made iOS changes and completed a build, and in DaVinci Resolve, where it reviewed video frame by frame. Those are demonstrations selected by Meta, not guarantees for every local machine or project.

Access is broad from day one. The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference is offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The model also supports several inference engines, including vLLM, SGLang, Transformers, llama.cpp and MLX, giving makers more ways to test it.

The release still needs independent scrutiny. Meta reports benchmark results and says the model met its own safety thresholds, while a full evaluation report is due later. SiliconANGLE reports that Meta’s scores are strong on some agent and tool-use tests but not uniformly ahead of Chinese open models. The weights use Meta’s Muse Model License with limited commercial terms, so “open-weight” does not mean unrestricted commercial use.

For developers, the practical news is a serious new option for local or hosted agent experiments. Running it locally requires exceptional hardware, and the license, speed, energy use and actual task quality still need checking. For the wider market, Meta is making a frontier-scale model more portable while Apple’s hardware becomes part of the story. That could expand experimentation, but it does not remove the need for permissions, isolation, logging and human review when an agent can change software or handle files.