Experimental Google models

This page details the experimental Google models that are available for testing in Vertex AI. It also shows you how to test the models by using the Vertex AI Studio and the Vertex AI API. Experimental models don't follow Google's standard model lifecycle plan and versioning scheme, and one model can be swapped for another without prior notice. We also don't guarantee that an experimental model will become a stable model in the future.

Test experimental models

Select one of the following tabs for instructions on how to test experimental Google models.

Console

To test an experimental Google model in Vertex AI Studio, do the following:

  1. In the Google Cloud console, go to Vertex AI Studio.

    Go to Vertex AI Studio

  2. In the Generate with Gemini section, click Open freeform.

  3. In the Model drop-down, select the model ID of an experimental Google model that you want to test.

    For a list of Google experimental model IDs, see Current experimental models on this page.

  4. In the Prompt field, enter your prompt.

  5. Click Submit.

REST

You can send API requests to experimental Google models by specifying its model ID in the endpoint. For example, to send a request to gemini-pro-experimental, the endpoint would look like the following:

https://us-central1-aiplatform.googleapis.com/v1/projects/my-project/locations/us-central1/publishers/google/models/gemini-pro-experimental:streamGenerateContent

For detailed instructions on how to use the API, see Multiturn chat.

Current experimental models

The following table gives you the details of the current experimental Google models that are available for testing:

Model ID Model family Date deployed Model characteristics/features
gemini-pro-experimental Gemini 1.5 Pro August 27, 2024

Better performance on coding and complex prompts compared to the previous experimental model.

gemini-flash-experimental Gemini 1.5 Flash August 27, 2024

General feedback-based improvements from the previous experimental model.

Previous experimental models

The following table gives you the details of the previous experimental Google models:

Specification Details
Model ID gemini-experimental
Date deployed July 16, 2024
Model family Gemini 1.5
Model characteristics/features

General feedback-based improvements from the previous experimental model.

Pricing

You aren't billed for the usage of experimental Google models.

Data privacy and feedback

Google does not log your input or output data, and we need your feedback to help improve our experimental models. To provide feedback, fill out the feedback form.

Quota and limitations

Experimental Google models have a quota of 10 queries per minute (QPM). This quota can't be increased.

Experimental Google models don't support the configuration of advanced parameters or the region, which is set to us-central1.

What's next