NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI · #286 most downloaded on PyPI
Vertex AI API client library
Last release 4 days ago
16 Sep 2026
Ships on a steady schedule
a new release about every 2 weeks
Nearly every release is documented
notes for 60 of the last 60 stable releases
2 versions withdrawn
withdrawn after publishing
6 years old
221 releases · first in 2020
One column per quarter.
Add _ModelBasedMetric base class to vertexai.preview.evaluation.metrics and allow metric spec customization
vertexai.preview.evaluation.metrics and allow metric spec customization (536f1d5)_AutomaticMetric type and support customizing metric spec for automatic pointwise metrics. Add Rouge metric to vertexai.preview.evaluation.metrics (f6b6dee)vertexai.preview.evaluation.metrics (4742a87).google.cloud.aiplatform.v1beta1.content.Candidate.FinishReason (5a6ce78)FeatureGroup.create() - remove unused entity id column param (c46f3e9)distibution in message .google.cloud.aiplatform.v1beta1.model_monitoring_stats.ModelMonitoringStatsDataPoint is changed. (5a6ce78)name in message .google.cloud.aiplatform.v1beta1.cached_content.CachedContent is changed (5a6ce78)source in message .google.cloud.aiplatform.v1beta1.tool.Retrieval is added. (5a6ce78)time_series_dataset and video_dataset classes (2cf4cb9)Add retry_timeout to EvalTask in vertexai.preview.evaluation
retry_timeout to EvalTask in vertexai.preview.evaluation (4d9ee9d)time_series_dataset and video_dataset classes (d5dc7b5)view model evaluations when AustoSxS pipeline completes
_append_tool_name list index out of range
Remove Vertex SDK data science package
Add a cloneable protocol for Reasoning Engine.
cloneable protocol for Reasoning Engine. (8960a80)GenerationConfig.response_schema field (#3772) (5436d88)response_schema parameter to the GenerationConfig class (b5e2c02)seed parameter to the TextGenerationModel's predict methods (cb2f4aa)Add support for ToolConfig in the LangChain template
BatchPredictionJob.submit method (4d091c6)Add DeprecationWarning to vertexai.preview predictive models SDK
vision_models.Image._mime_type property to make vision_models.Image compatible with generative_models.Image (6557d88)adapter_size parameter (88188d2)output_dimensionality and learning_rate_multiplier parameters to text embedding tuning (Preview only) (cc8bc96)AutomaticFunctionCallingResponder (58e6ac9)Add deprecation warnings when using Ray v2.4
Candidate.grounding_metadata property (b22a8b8)_TunableTextEmbeddingModelMixin.get_tuned_model method (eb05ac4)Register TensorFlow models from Ray checkpoints for more recent TensorFlow version, addressing the deprecation of SavedModel format in keras 3
GenerationConfig (0599ca1)MatchingEngineIndexConfig. (c0e7acc)Part.mime_type property (819a44e)Add support for reading requirements from a file.
GenerationResponse.prompt_feedback property (efd5a72)GenerationResponse.usage_metadata property (0654c35)NOT_EQUAL for MatchingEngineIndexEndpoint numeric_restricts. (aa918e3)Add display experiment button for tuning in Ipython environments
Experiment.dashboard_url property (c8eec21)SafetySetting.method (probability or severity) (317ab8f)text-bison@002 (1f27c3e)GenerativeModel's handling of tuned models from different region (bf33fb3)GenAI - Automatic Function Calling feature
Add an arg to turn off Ray metrics collection during cluster creation
accelerator_type in RLHF tuning (e51c977)Add custom tool context manager for telemetry
function_calls shortcut property to Candidate class. (d6490ff)max_context_length tuning parameter (e5daae9)shard_size for MatchingEngineIndex create index. (6dbf7d3)get_tuned_model deploying the model every time (e1c7870)from_dict methods for content types (GenerationResponse, Candidate, Content, Part) (613ce69)storage.googleapis.com URL import in vision_models.Image (2690e72)Add update_mask to MatchingEngineIndex upsert_datapoints() to support dynamic metadata update.
update_mask to MatchingEngineIndex upsert_datapoints() to support dynamic metadata update. (81f6a25)CodeGenerationModel (code-bison) (fbf2f7c)timeout for MatchingEngineIndex and MatchingEngineIndexEndpoint APIs. (09d1946)VideoSegmentConfig in Multimodal Embeddings API call (665e78c)Handle case when no metadata is returned from model.predict
Add Candidate.grounding_metadata
to_dict() methods to response and content classes (a78748e)GenerativeModel.start_chat(response_validation: bool = True) parameter (94f7cd9)text-bison@002, update pipeline template version and prediction row limit for pipeline-based LLM Evaluation SDK (35e57b3)imagetext, imagegeneration) (4109ea8)MatchingEngineIndexEndpoint query response. (83cb52d)Add numeric_filter to MatchingEngineIndexEndpoint match() and find_neighbor() private endpoint queries.
numeric_filter to MatchingEngineIndexEndpoint match() and find_neighbor() private endpoint queries. (679646a)BatchPredictionJob.submit method - a non-blocking version of BatchPredictionJob.create (b62ddcd)Fixed the PipelineJob bucket creation after a breaking change in Google Cloud Storage client library (a8b01e0), closes #2936
return_full_datapoint for MatchEngineIndexEndpoint match(). (ad8d9c1)MatchingEngineIndexEndpoint find_neighbors() public endpoint query. (42c7e08)low_level_batch_size in match() for MatchingEngineIndexEndpoint private query. (67e593b)MultiModalEmbeddingModel.get_embeddings(dimension=...) parameter (1d9bd23)MatchingEngineIndex create index. (a00db07)MatchingEngineIndexEndpoint match() and read_index_datapoints(). (61cff4b)Pin google-cloud-aiplatform[tensorboard] dependency on tensorflow < 2.15.0 due to breaking change introduced in tensorboard 2.15.1
fraction_leaf_nodes_to_search_override. Add support for private endpoint in find_neighbors. (cd31c13)Adding serving_container_grpc_ports parameter to Model.upload() method
Release GenerativeModel support for Gemini
logprobs, presence_penalty, frequency_penalty, and logit_bias generation parameters (1449344)read_index_datapoints for private network. (c9f7119)read_index_endpoint private endpoint support. (3d8835e)Add additional parameters to Model.upload().
INTEGER and BOOL parameters are not cast to the correct type. (9a204c4)Add numeric_restricts to MatchingEngineIndex find_neighbors() for querying
numeric_restricts to MatchingEngineIndex find_neighbors() for querying (6c1f2cc)remove_datapoints() to MatchingEngineIndex. (b86a404)upsert_datapoints() to MatchingEngineIndex to support streaming update index. (7ca484d)encryption_spec_key_name for MatchingEngineIndexEndpoint create. (7740132)encryption_spec_key_name in MatchingEngineIndex create_tree_ah_index and create_brute_force_index (595b580)Add option to not use default tensorboard
Add encryption_spec_key_name to MatchingEngineIndex create_tree_ah_index and
encryption_spec_key_name to MatchingEngineIndex create_tree_ah_index and (1a9e36f)encryption_spec_key_name, enable_private_service_connect,project_allowlist to MatchingEngineIndexEndpoint create. (750e17b)index_update_method to MatchingEngineIndex create() (dcb6205)disable_attribution grounding parameter (91e985a)ChatModel, CodeChatModel) (755c3f9)ImageGenerationModel (791eff5)tabular_dataset class (6f40f1b)Add per_crowding_attribute_neighbor_count, approx_num_neighbors, fraction_leaf_nodes_to_search_override, and return_full_datapoint to MatchingEngineIn
per_crowding_attribute_neighbor_count, approx_num_neighbors, fraction_leaf_nodes_to_search_override, and return_full_datapoint to MatchingEngineIndexEndpoint find_neighbors (33c551e)per_crowding_attribute_num_neighbors approx_num_neighborsto MatchingEngineIndexEndpoint match() (e5c20c3)per_crowding_attribute_num_neighbors approx_num_neighborsto MatchingEngineIndexEndpoint match() (53d31b5)per_crowding_attribute_num_neighbors approx_num_neighborsto MatchingEngineIndexEndpoint match() (4e357d5)enable_checkpoint_selection tuning evaluation parameter (eaf4420)*-bison-32k models (9eba18f)CodeChatModel tuning to GA (621af52)Add preview count_tokens method to CodeGenerationModel
count_tokens support to ChatModel (preview) (01989b1)CodeGenerationModel tuning to GA (87dfe40)accelerator_type in tuning (98ab2f9)ImageGenerationModel (cbe3a0d)Add serializer.register_custom_command()
Add Model Garden support to vertexai.preview.from_pretrained
count_tokens method to the preview TextGenerationModel and TextEmbeddingModel classes (6a2f2aa)ChatModel tuning to GA (7d667f9)Lightning trainer fails to be unwrapped in remote training
Add Custom Job support to from_pretrained
Endpoint class (e9eb159)context (f7feeca)LLM - Added stop_sequences parameter to streaming methods and CodeChatModel
stop_sequences parameter to streaming methods and CodeChatModel (d62bb1b)CodeGenerationModel now supports safety attributes (c2c8a5e)TextEmbeddingInput.task_type parameter. (2e3090b)fix: LLM - De-hardcoded the max_output_tokens default value for the CodeGenerationModel
Add disable_retries option to custom jobs.
stop_sequences in inference (6f7ea84)TextGenerationResponse.raw_prediction_response (f8f2b9c)TextGenerationModel tuning to GA (62ff30d)TextEmbeddingInput), auto_truncate parameter and result statistics (cbf9b6e)ImageCaptioningModel to GA (7575046)ImageQnAModel to GA (fd5cb02)MultiModalEmbeddingModel to GA (e99f366)width and height parameters from ImageGenerationModel.generate_images since the service has dropped support for image sizes and aspect ratios (52897e6)LLM - Added tuning support for chat-bison models
Add model.evaluate() method to Model class
Add preview CustomJob which can be run on persistent resource
textembedding models (preview) (a368538)imagetext model) and multi modal embedding (multimodelembedding model) (preview) (9bbf1ea)get_tuned_model for the future models that are not text-bison (1adf72b)LLM - Released the BatchPrediction to GA for TextGenerationModel
LLM - Released the Chat models to GA
Add sdk support for xai example-based explanations
Add submit for CustomTrainingJob and CustomContainerTrainingJob which won't block until complete.
learning_rate in tuning (c6cdd10)ChatMessage class (7bf7634)Add additional scheduled pipelines client getters and unit tests.
Add additional scheduled pipelines client create method unit tests.
1.26.0 (2023-06-07)
enable_access_logging parameter to Endpoint.deploy() method, minor edit to batch_predict() docstring (794cedd)BatchPredictionJob.create (13b11c6)ChatModel.start_chat being ignored (a0d815d)_construct_sdk_resource_from_gapic should use the project from the resource name instead of the default project. (162b2f2)Add support for Large Language Models
Add preview capability to deploy models with shared resources.
Add ExperimentRun.get_logged_custom_jobs method
Implement Model.copy functionality.
Add support for enable_dashboard_access field for Training jobs in SDK
Add a return value (ClassificationMetrics) for the log_classification_metrics()
Add default skew threshold to be an optional input at _SkewDetectionConfig and also mark the target_field and data_source of skew config to optional.
Endpoint.undeploy_all() doesn't undeploy all models (9fb24d7)Adds the temporal fusion transformer (TFT) forecasting job
Add explanationSpec to TrainingPipeline-based custom jobs
Add Feature Store: Streaming Ingestion (write_feature_values()) and introduce Preview namespace to Vertex SDK
Add a sample for get_experiment_run_artifacts
Added proto message conversion to MDMJob.update fields
## 1.18.1 (2022-10-10) ### Bug Fixes * deps: Allow protobuf 3.19.5
Add deleteFeatureValues in aiplatform v1beta1 featurestore_service.proto
Add enable_simple_view to PipelineJob.list()
Your coding agent can read these notes before it upgrades. Set up the MCP server →