Muse Spark 1.3

Muse Spark 1.3 from Meta is a reasoning model trained for agentic workflows and competitive coding. It supports built-in tool calling with Model Context Protocol (MCP) support, accepts text and image inputs, and offers a 1M-token context window for multi-step tasks.

Requests on the global /locations/global/publishers/meta/v1/responses endpoint pass the bare model ID muse-spark-1.3 in the model field. For an example request, see Use Meta models.

View model card in Model Garden

Model ID meta/muse-spark-1.3
Launch stage Preview
Supported inputs & outputs
  • Inputs:
    Text, Image
  • Outputs:
    Text
Token limits
  • Maximum input tokens: 1,048,576
Capabilities
Usage types
Versions
  • meta/muse-spark-1.3
    • Launch stage: Preview
    • Release date: September 24, 2026
Supported regions

Model availability

  • Global
    • global endpoint
Quota limits

global endpoint:

  • Context length: 1,048,576

Pricing See Pricing.

Inference parameters and limitations

Keep the following inference parameters and limitations in mind when sending requests:

  • Reasoning effort: Set reasoning.effort to minimal, low, medium, high, or xhigh. max isn't available on Google Cloud, and none isn't supported. Pricing is the same at every effort level.
  • Reasoning output: Reasoning content is returned in encrypted form for multi-turn replay. Optional reasoning summaries (reasoning.summary) aren't returned on every response.
  • Tool calling: Supports built-in tool calling with Model Context Protocol (MCP) support. tool_choice supports auto only (none, required, and named-function choices aren't supported).
  • Recommended sampling: temperature 1.0, top_p 1.0 (top_k disabled).
  • Known limitations: Audio, video, and PDF inputs, chat completions, and stateful response storage (store: true) aren't supported. The model isn't trained for tabular input, so results on tabular data can vary. For the full capability list, see the table earlier on this page.

What's next