Meta Muse Models on Google Cloud: What Ember, Spark and Wonder Support in Preview

Meta Muse models on Google Cloud are available in preview through the global endpoint of Google’s Gemini Enterprise Agent Platform. The lineup includes Muse Ember 1.0, Muse Spark 1.3 and Muse Wonder 1.0. For a team testing them through Model Garden, the quickest way to narrow the list is to check the input formats required by its workflow.

The Google Cloud release notes describe access through the OpenAI-compatible Responses API. Google’s model documentation distinguishes the inputs supported on that particular service, which matters when evaluating claims about what the broader Muse family can handle.

Match each model to the files you send

The Muse Ember model card lists text and image input with text output. Google positions Ember for repeated agent tasks such as extraction, routing and tool calls. The card lists a maximum input length of 1,048,576 tokens and marks structured output and function calling as preview capabilities.

Google’s cloud release notes list text, image and PDF support for Wonder and Spark 1.3. If your application sends complete PDFs, start by testing those two on a small set of representative files. If you already extract text or send page images, you can also evaluate Ember using that same preprocessing pipeline.

The Wonder model card focuses on complex software engineering and longer agent tasks. The Spark 1.3 documentation describes a reasoning model designed for agent workflows and coding. Use those descriptions to select an initial test, then compare the actual outputs against the requirements of your application.

Check the endpoint and model ID before porting code

Google’s Meta integration instructions describe enabling a model through its Model Garden card. Requests use the Meta publisher Responses endpoint for your Google Cloud project. The request body expects the bare identifier, such as muse-ember-1.0, muse-spark-1.3 or muse-wonder-1.0.

Start with a basic text request and confirm that your project can access the model. Then add one feature at a time. For an application that calls tools, first verify that a simple call returns the expected arguments, then validate those arguments in your application before executing anything. That makes it easier to tell a permissions problem from a mismatch in the request body.

Run a workload test before deciding

Choose a small sample that includes an ordinary case and a troublesome one. A document extraction test might include a clean invoice, a scanned page with a faint total and a record missing one of your required fields. Track which values the model recovered and which ones it guessed. For coding, use an isolated repository task with a reproducible failure and a test you can rerun.

Log input size, elapsed time, output validity and the number of retries. Evaluate the cost of producing a usable result for the full task, including any preprocessing and extra model calls. A large context window gives you room to submit more material, but accuracy on the relevant fields still needs measurement.

Before putting a preview model in a workflow people depend on, check its current pricing, quota and support in the Google Cloud console. A small comparison using the same inputs gives your team concrete evidence for choosing which Muse model to continue testing.

For multi-step workflows, test the planner and the worker separately before comparing a combined run. A planner can propose the right sequence while the worker still extracts the wrong value from a file. Write down the expected result at each stage and log which model produced it. If you replace one stage at a time, you can see whether a change helps accuracy, speed or both. Keeping the workflow small at first also makes repeated measurements cheaper and easier to interpret.

Illustrative photograph.