Muse Spark Open

Muse Spark Open is the open-weight release Meta kept promising through the Muse Spark 1.2 and 1.3 launches without naming a date, a variant, or a license, and on October 15, 2026 all three arrived at once. It is a sparse mixture-of-experts with 816 billion total parameters and 42 billion active per token, natively multimodal, with a 1 million token context window, published under the Muse Model License rather than Apache 2.0 or plain MIT: Meta describes the commercial terms as limited but free for many developers, so read the license before you ship rather than assuming the Llama posture carried over. Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, and hosted inference runs through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The hosted rate matches the Muse Spark Standard tier exactly, $1.25 per million input tokens and $4.25 output with cached input at $0.15, which is the detail that makes the release coherent: Meta is not undercutting its own API, it is letting you leave it. The hardware story is the one that travelled furthest. Meta worked with Apple so the model runs on two 512GB M5 Ultra Mac Studios, and reporting puts the practical floor at roughly 1TB of unified memory, which is a workstation pair rather than a rack and is unusual for a model this size. Meta calls it the most capable open-weight model from a US lab and says its own evaluations beat other popular open-weight models; the launch post publishes its benchmark section as a chart rather than a table, so no row-level figures are carried here. Note what Meta did not do: it has not put Muse Spark Open on the Meta Model API, so the first-party endpoint still serves 1.3 and the open model is served by everyone else.