Muse Spark Open

Muse Spark Open is Meta's open-weight multimodal mixture-of-experts model for coding and agentic work. Meta says the release has 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. The weights are available through the official Hugging Face collection, while hosted access is available from several API platforms. This is a distinct deployment option from the hosted Muse products: a team can evaluate the weights and choose its own runtime, infrastructure, and data boundary.

What it is for

Use it when a code or tool-using agent needs a very large multimodal model but the deployment path must remain under your control. Meta specifically positions Muse Spark Open for long-horizon tasks, complex software engineering, multi-step reasoning, and coding. The official release describes two Apple M5 Ultra Mac Studios with MLX as a desktop path for this model class. That is a capability demonstration, not a general hardware recommendation: its memory, quantization, runtime, concurrency, and latency requirements must be measured on the intended setup.

Key capabilities

• Open weights: Meta released the weights through its official Hugging Face collection.
• Sparse multimodal model: 816B total parameters with 42B active parameters per the official announcement.
• Long context: A one-million-token context window for large repositories and multi-step agent traces.
• Deployment choice: Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI as hosted availability paths.
• Post-training paths: Meta's build guide lists TorchTitan, SGLang with Miles, NVIDIA NeMo AutoModel, Hugging Face TRL, and managed APIs from Fireworks AI and Together AI.

A practical evaluation workflow

Start with a small, non-production coding or document task set. Run the official model card's recommended format in one runtime, keep tool permissions read-only, and record context length, time to first token, tool-call validity, recovery after failed calls, and end-to-end cost. Then repeat the same suite through one hosted provider. Compare the results with Muse Spark 1.2 or a smaller local model before granting write, shell, or network actions.

Limits and alternatives

Open weights do not make deployment inexpensive or simple. The official release's desktop example uses two high-end systems, and the Hugging Face collection is gated. Provider availability, price, model formats, and terms can change, so verify them at the point of integration. For a hosted terminal workflow rather than model operations, compare Muse Code or Codex CLI.

Getting started

Read Meta's official announcement, request access to the official Hugging Face collection, and benchmark one constrained workflow before connecting the model to production tools.