Muse Spark Open

What the open-weight release contains

The official checkpoint is meta-models/Muse-Spark-Open-816B-A42B, a sparse mixture-of-experts model with a 1,048,576-token inference context. The 42B active count describes computation per token; the deployment must still accommodate the resident weights and runtime state. Its multimodal inputs produce text, not native image or video output.

Commercial use has license conditions

The Muse Model License permits use, modification and distribution under its terms. Products using the software must prominently display Built with Muse Spark Open, and distribution must retain the required notices, license and Usage Policy.

• Read the complete license and Usage Policy

  Use the actual terms for your intended deployment and distribution.

• A separate agreement is required when the licensee or affiliates offer Model as a Service and their aggregate gross revenue exceeds USD20 million in any consecutive twelve-month period.
• The threshold concerns total licensee/affiliate revenue, not just this model's API revenue. The license explicitly excludes internal use, research and non-MaaS commercial uses, including qualifying features in the licensee's own product.
• The Usage Policy remains applicable. Downloadable weights do not mean unrestricted open-source terms.

Local memory depends on format and runtime

Meta documents approximately 537GB of NVFP4 checkpoint files and 1.65TB for BF16. These file sizes are different from the model card's stated runtime budgets of at least 700GB VRAM for NVFP4 and 2TB for BF16. KV cache, activations, context length and serving configuration also consume memory.

• Compare documented hardware platforms

  Check the exact runtime, format and interconnect before provisioning.

• The official Mac guide specifies NVFP4 with MLX on two M5 Ultra Mac Studios with 512GB each, or four with 256GB each.
• Two Macs are one supported configuration, not a universal requirement. NVIDIA, AMD and Intel recipes have their own hardware and format constraints.
• Alternative quantizations require their own memory and quality checks; an advertised checkpoint size is not a measured speed or a guarantee that a full-context workload fits.

Hosted access is separate from Meta Model API

Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as inference partners. The developer overview explicitly excludes Muse Spark Open from Meta Model API; do not reuse a muse-spark-1.3 request and assume it selects the open-weight checkpoint. Account, region and feature availability belong to the selected provider.

• Muse Spark Open API record

  See the documented OpenRouter ID, prices, limits and an example request.

• Muse Spark 1.3

  Compare the separate first-party hosted model and its reasoning settings.

Check the routed provider's token limits

On October 17, OpenRouter lists meta/muse-spark-open at USD1.25 per million input tokens and USD4.25 per million output tokens across five endpoints. Four list USD0.15 cached input; Nebius does not list that field. Modal advertises 1,000,000 context tokens and 262,144 maximum completion tokens. Four other endpoints advertise 1,048,576 and 943,718, while the public model page lists 262,144 completion tokens. These are different source scopes, not one reconciled guarantee.

• Inspect current provider endpoints

  Pin a provider when its contract matters and verify the actual request budget.

Tool use needs a controlled runtime

Meta's prompting guide documents conversation roles, tool calls and separate reasoning output. Muse Spark Open always reasons; native reasoning_effort accepts low, medium, high or xhigh, with xhigh as the default. This differs from Spark 1.3's settings. A tool call still needs an external executor, permissions, returned tool results and checks of the final work.

• Start with a repository task or supplied document with a reference answer and explicit completion criteria.
• Review tool permissions, failed steps, output correctness and billed usage before allowing a longer workflow.
• Self-hosting gives infrastructure control; it does not make an agent reliable or remove the operator's deployment responsibilities.

Read launch evaluations with their conditions

Meta's xhigh model-card table reports 89.2 on MCP-Atlas, 60 on DeepSWE 1.1 and 76.4 on Terminal-Bench 2.1 with the listed harnesses. The same table includes stronger competing results on several coding and general-intelligence rows. The announcement's strongest-US-open-weight positioning is Meta's claim, not a universal ranking.

• Meta describes on-device Muse Code work on the Ice Cubes iOS app and a DaVinci Resolve editing workflow. These are publisher workflow reports, not independent runs or proof of native video generation.
• For adoption, compare the same input, tools and success criteria, then record retries, total cost and review effort.

Choose between Open, Spark 1.3 and Wonder

Choose Open when downloadable weights, customization or infrastructure control justify the operational and license work. Spark 1.3 is a separate Meta Model API and Muse Code offering. Wonder is another hosted model family; its access, price and Full subscription variant do not transfer to Open.

• Muse Wonder

  Check the separate model, privacy modes and subscription-only Full boundary.

• Kimi K3

  Another open-weight model with its own license, deployment scale and API contract.

Useful first jobs

• Evaluate a customizable coding model
• Build reviewed multimodal tool workflows
• Compare hosted inference with controlled self-hosting