Muse Spark Open

Muse Spark Open is Meta's October 15, 2026 open-weight multimodal model: 816B total parameters, 42B active per token and text output.

Muse Spark Open explained
License, hardware choices, release context and evaluation limits.
Official weights and license
Check the downloadable model card and access conditions.
OpenRouter hosted offer
Check the provider selected for your API request.

Open WeightsCodingTool UseMultimodalMeta

Self-hosted BF16/NVFP4 or third-party hosted inference; large memory and runtime-specific requirements

OpenRouter endpoints checked October 17: $1.25/M input, $4.25/M output; $0.15/M cache read on four endpoints, not listed on Nebius. No universal Meta API rate is implied.

Scoped Terms

🚀Function Overview

The official checkpoint accepts text, image, video and audio input with text output and 1,048,576-token model context. Try it now and see the results!

Key Features

816B total / 42B active sparse mixture of experts; BF16 and NVFP4 checkpoints.
Hosted API ID: meta/muse-spark-open on OpenRouter, separate from the downloadable meta-models/Muse-Spark-Open-816B-A42B identifier.
October 17 endpoint listing: Modal advertises 1,000,000 context tokens / 262,144 maximum completion tokens.
Together, Nebius, Fireworks and Baseten endpoints advertise 1,048,576 context / 943,718 completion; the public OpenRouter model page instead lists a 262,144 completion cap. Verify the selected endpoint before budgeting.
Native reasoning_effort supports low, medium, high and xhigh (default); reasoning cannot be disabled. Hosted controls and accepted values must follow the provider.
Tool calling and structured output support are provider-specific. Tools still need an external executor and reviewed permissions.
Meta's MLX/NVFP4 guide supports two 512GB or four 256GB M5 Ultra Mac Studios. GPU and other quantization paths have separate memory and runtime requirements.
Launch evaluations are publisher-reported and configuration-specific; no model execution or task-quality test was performed here.

Use Cases

•Evaluate self-hosted coding inference
•Build reviewed tool-calling agents
•Analyze multimodal input with a verified reference answer

⚙️Input Parameters

model

string

meta/muse-spark-open at https://openrouter.ai/api/v1/chat/completions. Authenticate with an OpenRouter API key; do not send this host ID to Meta Model API.

messages

array

OpenRouter Chat Completions role/content messages. The example is text-only. Follow the selected provider's media schema for images, audio, video and files.

max_tokens

integer

Explicit completion budget within the selected provider's context and output limits. Reasoning consumes budget; inspect finish_reason for truncation.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "meta/muse-spark-open",
  "messages": [
    {
      "role": "user",
      "content": "Review this loop for an off-by-one error and explain the correction: for (let i = 0; i <= items.length; i++) { consume(items[i]); }"
    }
  ],
  "max_tokens": 4096
}

Output Results

Illustrative OpenRouter request; not executed here. Inspect choices[].message.content, finish_reason and usage in the real response. This text-only example does not verify media support or task accuracy.

Quick Actions

Technical Specifications

Hardware Type: Self-hosted BF16/NVFP4 or third-party hosted inference; large memory and runtime-specific requirements
Commercial Use: Conditional Muse Model License: retain required notices and Usage Policy, display Built with Muse Spark Open, and check the separate-agreement condition for qualifying MaaS businesses above USD20M aggregate licensee/affiliate revenue in any consecutive twelve months. Provider service/privacy terms also apply.
Pricing: OpenRouter endpoints checked October 17: $1.25/M input, $4.25/M output; $0.15/M cache read on four endpoints, not listed on Nebius. No universal Meta API rate is implied.