What happened?
Meta Superintelligence Labs launched Muse Spark Open on October 15, 2026, described as the most capable US open-weight model to date. It is a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active per token, plus a 1-million-token context window.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering and multi-step reasoning. Weights are available on Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry under the Muse Model License.
Why is this important?
For the first time, a model of this scale from a US open-weight lab can run on desktop hardware: specifically on two Apple Mac Studios with the M5 Ultra chip using the MLX framework. This democratizes access to frontier AI for developers who prefer to run locally instead of relying solely on APIs.
The model accepts text, image, video and audio as input and is optimized for multi-agent orchestration, with a commentary channel that explains intentions before acting.
What changes in practice?
Developers can download and customize the model with PyTorch TorchTitan or MLX for Apple Silicon. It is available via APIs on platforms like Together AI, Fireworks, Groq, OpenRouter and others. The license allows commercial use with restrictions for high-revenue companies that offer the model as a service.
This positions Meta in the open-weight model race against Chinese competitors and others, offering a strong option for coding agents and agentic tasks without mandatory cloud inference costs. More details on the official Meta AI Research blog.
By GeekikiBot
