How to Use the Muse Spark 1.3 API

Muse Spark 1.3 is available on EmpirioLabs. Meta built this checkpoint for long-horizon work: agentic runs, multi-step coding, and tasks that stay in one thread long enough to accumulate context. It pairs a 1,048,576-token context window with image, video, and PDF understanding, function calling, strict structured output, built-in web search, and native Responses tools.

What Muse Spark 1.3 supports

Muse Spark 1.3 accepts text, up to 50 images, MP4 video, and PDF documents. PDF processing covers text from the first 100 pages and page images from the first 50 pages, which suits long reports, specifications, and visual documents.

Reasoning is always on. Use reasoning_effort from minimal through xhigh to control how much work the model spends before answering. Reasoning traces stay private, and reasoning tokens count as output tokens against the request output budget.

The long-context behaviour is the headline: Meta reports the model tracking detail across very long threads, asking for clarification on ambiguous prompts, and checking in before consequential actions rather than guessing.

How to call Muse Spark 1.3

Use the OpenAI-compatible chat completions endpoint and set model to muse-spark-1-3:

curl https://api.empiriolabs.ai/v1/chat/completions \
  -H 'Authorization: Bearer $EMPIRIOLABS_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "muse-spark-1-3",
    "reasoning_effort": "medium",
    "messages": [
      {
        "role": "user",
        "content": "Plan a staged migration of a busy service to a new database, and flag what could go wrong."
      }
    ]
  }'

The same model is available through the Responses, Messages, and Google-compatible content-generation endpoints. The Responses endpoint preserves custom tools and deferred tool discovery with tool_search. Streaming responses carry final usage data so token, cache, and tool settlement stays auditable.

Web search, tools, and structured output

Turn on built-in web search when an answer needs current information. Every executed search query is reported in usage.tool_usage and adds the catalog-listed per-search charge, and a single request can run several searches. Function tools can be called in parallel, and cited sources are appended when the model returns them.

For machine-readable output, use JSON mode or supply a strict JSON Schema. Schema mode is enforced, which makes it dependable for extraction, routing, evaluation, and agent workflows that need a stable response contract.

Pricing and data use

Muse Spark 1.3 uses pay-as-you-go token pricing, with a discounted rate for cached input and a separate per-query rate when built-in web search runs. Current rates are always on the model page and the pricing page. There is no subscription and no minimum.

EmpirioLabs serves the Standard checkpoint. Prompts and completions sent to it are not used by Meta to train its models.

Notes worth knowing before your first call

• Reasoning tokens are drawn from the same budget as your answer. A tight max_tokens can be consumed entirely by reasoning and return a length-truncated response with no visible content, so give long-horizon prompts a generous output budget.
• Unlike Muse Spark 1.1 and 1.2, this checkpoint does not accept audio input. Requests carrying audio are rejected immediately rather than silently ignored, so route speech to 1.2 or to a dedicated transcription model.
• tool_choice supports auto and none. Forcing a specific function by name is not available on this model.

Get started

Try Muse Spark 1.3 in the Playground, read the API reference, or see full specifications on the model page.