GenerativeModel

abstract @FeatureStatus int checkStatus()

Checks the current availability status of the content generation feature.

abstract void clearCaches()

Clears all caches created by prefix caching.

This experimental method clears all caches created by prefix caching. When promptPrefix is provided in generateContent or generateContentStream, the system caches its processing to reduce inference time for subsequent requests sharing the same prefix. This method clears all such created caches.

abstract void close()

Releases resources associated with the content generation engine.

This should be called when the GenerativeModel is no longer needed. Can be safely called multiple times.

abstract @NonNull CountTokensResponse countTokens(@NonNull GenerateContentRequest request)

Counts the number of tokens in the request.

The number of tokens counted includes input and output tokens. The result can be compared with getTokenLimit to check if the request is within the token limit.

abstract @NonNull Flow<@NonNull DownloadStatus> download()

Downloads the required model assets for the content generation feature if they are not already available.

Use this method to proactively download models before inference. The returned Flow emits DownloadStatus to report progress and completion status.

default @NonNull GenerateContentResponse generateContent(@NonNull String prompt)

Performs asynchronous content generation on the provided input prompt.

This is a convenience method that wraps the input prompt in a GenerateContentRequest with default generation parameters.

abstract @NonNull GenerateContentResponse generateContent(@NonNull GenerateContentRequest request)

Performs asynchronous content generation on the provided input request.

This is the standard, non-streaming version of inference. The full generation suggestions are returned once the model completes processing.

This method is non-blocking. To handle the result, callers should use try/catch to handle the returned GenerateContentResponse or a potential GenAiException.

The coroutine that runs generateContent is cancellable. If the inference is no longer needed (e.g., the user navigates away or input changes), the coroutine can be cancelled.

Note that inference requests may fail under certain conditions such as:

• Invalid or malformed input in the GenerateContentRequest.

• Exceeded usage quota or failed safety checks.

default @NonNull GenerateContentResponse generateContent(
    @NonNull String prompt,
    @NonNull StreamingCallback streamingCallback
)

Performs streaming content generation inference on the provided input prompt.

This is a convenience method that wraps the input prompt in a GenerateContentRequest with default generation parameters.

abstract @NonNull GenerateContentResponse generateContent(
    @NonNull GenerateContentRequest request,
    @NonNull StreamingCallback streamingCallback
)

Performs streaming content generation inference on the provided input request.

Partial results are delivered incrementally through the provided StreamingCallback. The function suspends until all results are received and returns the complete, final GenerateContentResponse. If streaming is interrupted by a GenAiException, consider removing any already streamed partial output from the UI.

The coroutine that runs generateContent is cancellable. If the inference is no longer needed (e.g., the user navigates away or input changes), the coroutine can be cancelled.

Note that inference requests may fail under certain conditions such as:

• Invalid or malformed input in the GenerateContentRequest.

• Exceeded usage quota or failed safety checks.

default @NonNull Flow<@NonNull GenerateContentResponse> generateContentStream(@NonNull String prompt)

Performs streaming content generation inference on the provided input prompt.

This is a convenience method that wraps the input prompt in a GenerateContentRequest with default generation parameters.

abstract @NonNull Flow<@NonNull GenerateContentResponse> generateContentStream(@NonNull GenerateContentRequest request)

Performs streaming content generation inference on the provided input request.

Partial results are delivered incrementally through the returned Flow. Each GenerateContentResponse contains a single Candidate. The last emitted value contains a Candidate with a FinishReason that is a non-null FinishReason (e.g., FinishReason.STOP or FinishReason.MAX_TOKENS)

This streaming mode is useful to build a more responsive UI. Streaming can be interrupted by a GenAiException. In that case, consider removing any already streamed partial output from the UI.

The coroutine collecting the Flow is cancellable. If the inference is no longer needed (e.g., the user navigates away or input changes), the coroutine can be cancelled

Note that inference requests may fail under certain conditions such as:

• Invalid or malformed input in the GenerateContentRequest.

• Exceeded usage quota or failed safety checks.

Important: This function currently only supports a candidateCount of 1 in the GenerateContentRequest. Providing a candidateCount greater than 1 will result in an IllegalArgumentException. If you need to receive multiple candidates in the final result, please use the generateContent(request: GenerateContentRequest, streamingCallback: StreamingCallback) method instead.

abstract @NonNull String getBaseModelName()

Returns the name of the base model used by this generator instance.

The model name may be used for logging, debugging, or feature gating purposes.

abstract int getTokenLimit()

Returns total token limit for the API including both input and output tokens.

This limit can be used with countTokens to check if a request is within limits before running inference. The input size returned by countTokens plusing the output size specified by GenerateContentRequest.maxOutputTokens should be no larger than the limit returned by this method.

abstract void warmup()

Warms up the inference engine for use by loading necessary models and initializing runtime components.

While calling this method is optional, we recommend invoking it well before the first inference call to reduce the latency of the initial inference.