Muse models from Meta

Muse models from Meta are available for use as managed APIs on Agent Platform. You can stream your responses to reduce the end-user latency perception. A streamed response uses server-sent events (SSE) to incrementally stream the response.

Managed Meta models

The following models are available from Meta to use in Agent Platform. To find and enable a Meta model, go to its model card in Model Garden.

Muse Wonder from Meta

Muse Wonder from Meta is a high-capability reasoning model built for complex software engineering and long-horizon agentic work. In multi-step coding, autonomous agent, and computer-use workflows, it plans tasks, delegates to sub-agents, calls tools in parallel, and recovers from errors.

Muse Ember from Meta

Muse Ember from Meta is an efficient reasoning model built for high-volume agent tasks, including tool calls, coding subtasks, routing, extraction, and long-document processing. Pair it with Muse Wonder or Muse Spark 1.3 for multi-step planning and reasoning.

Muse Spark 1.3 from Meta

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 and a 1M-token context window for multi-step tasks.

Use Meta models

Muse models from Meta are served through the stateless OpenAI-compatible Responses API on the global endpoint. Chat completions and stateful response storage (store: true) aren't supported. Send POST requests to https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/meta/v1/responses and pass the bare model ID (muse-wonder-1.0, muse-ember-1.0, or muse-spark-1.3) in the model field.

Keep the following parameter details in mind when sending requests:

  • Inputs: Text and image inputs are supported. Audio, video, and PDF inputs aren't supported.
  • Reasoning effort: Set reasoning.effort to minimal, low, medium, high, or xhigh (Muse Ember also supports max). none isn't supported.
  • Tool calling: tool_choice supports auto only (none, required, and named-function choices aren't supported).

The following examples send a prompt request to Muse Wonder on the global endpoint. Replace PROJECT_ID with your Google Cloud project ID.

REST (curl)

curl -X POST \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  "https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/meta/v1/responses" \
  -d '{
    "model": "muse-wonder-1.0",
    "reasoning": {"effort": "high"},
    "input": "Write a Python binary search function with type annotations.",
    "max_output_tokens": 4096,
    "stream": false
  }'

Python (OpenAI SDK)

Install the openai and google-auth libraries (pip install --upgrade openai google-auth), set up Application Default Credentials (gcloud auth application-default login), and then send a request:

import google.auth
import google.auth.transport.requests
from openai import OpenAI

PROJECT_ID = "PROJECT_ID"
LOCATION = "global"

credentials, _ = google.auth.default(
    scopes=["https://www.googleapis.com/auth/cloud-platform"]
)
credentials.refresh(google.auth.transport.requests.Request())

client = OpenAI(
    base_url=f"https://aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/{LOCATION}/publishers/meta/v1",
    api_key=credentials.token,
)

response = client.responses.create(
    model="muse-wonder-1.0",
    reasoning={"effort": "high"},
    input="Write a Python binary search function with type annotations.",
    max_output_tokens=4096,
)
print(response.output_text)