Meta ships Muse Spark Open, an 8,160億-parameter open model

Muse Spark Open arrives as a native multimodal model rather than a text-only one, and Meta positions it for software engineering, multi-step logical reasoning and long-running autonomous agents — the workloads where context length and tool use matter most. The 1-million-token context window is the headline spec for that use case, letting an agent hold far more code and task history in view at once.

Distribution is split deliberately: hosted access through Together AI, Fireworks AI and Groq for teams that want an API, and downloadable weights via Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models for those who want to self-host. Meta's Apple Silicon optimization, using the MLX framework, goes further, supporting a setup of two connected Mac Studios with Apple M5 Ultra chips.

The local tests Meta describes are unusually concrete for a launch announcement: Muse Code was used to make iOS code changes — bug fixes, new features and a full build — to Ice Cubes, an open-source Mastodon client, and to drive DaVinci Resolve through frame-by-frame video review and vertical trimming. All of it ran on the local environment, which suggests Meta is pitching on-device agent work as a practical option rather than a demo.