How to Use the Muse Spark 1.1 API

Muse Spark 1.1 is Meta AI's frontier reasoning model, now live on EmpirioLabs. It pairs a 1,048,576-token context window with always-on reasoning, image and video understanding, function calling, strict JSON Schema structured output, and built-in web search that cites its sources. It is built for agentic tool work, coding, document and media analysis, and grounded research.

You can try it in the Playground or call it through the OpenAI-compatible API today. Full parameter details live on the docs page.

Pricing

Muse Spark 1.1 is pay-as-you-go with per-token billing. Prompt tokens, generated tokens, and cached input tokens each have their own rate, and cached input is billed at a discount whenever the service detects a repeated prompt prefix, which happens automatically with no cache configuration on your side. Built-in web search adds a small per-search-query fee only when a search actually runs. Current rates are always on the model page and the pricing page.

Quickstart

from openai import OpenAI

client = OpenAI(
    base_url="https://api.empiriolabs.ai/v1",
    api_key="YOUR_EMPIRIOLABS_API_KEY",
)

response = client.chat.completions.create(
    model="muse-spark-1-1",
    reasoning_effort="medium",
    max_tokens=16384,
    messages=[{"role": "user", "content": "Summarize the tradeoffs between SQL and NoSQL in five bullets."}],
)
print(response.choices[0].message.content)

The same model id works on /v1/chat/completions, /v1/responses, Anthropic-style /v1/messages, and the Google-compatible generateContent endpoint, so most existing SDKs can point at EmpirioLabs without code changes.

Built-in web search

Set "tool_web_search": true and the model can search the web on its own, ground the answer in what it finds, and append a Sources list with links when citations are available. Each executed search query adds the per-query fee shown on the model page, and searches that never run cost nothing.

Image and video input

Send images with standard image_url content parts and video with video_url parts, as an https URL or a base64 data URL. Multiple images in one request are supported. Both work together with web search when you need grounded answers about an image or clip.

Operational notes

• Reasoning is always on and cannot be disabled, and the model does not return its reasoning text, only the final answer. Reasoning tokens bill as output tokens and count against max_tokens, so give the model room: very small output caps can be consumed by reasoning before the visible answer appears. Use reasoning_effort (minimal, low, medium, high, xhigh) to control how hard it thinks.
• Supported sampling controls are temperature, top_p, seed, presence_penalty, and frequency_penalty. Stop sequences, top_k, logprobs, and multiple choices per request are not supported by this model.
• tool_choice supports auto and none. Structured output supports both JSON mode and strict JSON Schema.

Get an API key from the dashboard and you are one request away.