IBM announces the general availability of watsonx.ai version 2.3 (December 15, 2025). Watsonx.ai is an enterprise-grade AI development studio that runs on an open and trusted hybrid cloud infrastructure and helps developers operationalize and scale the development of AI applications by bringing together traditional machine learning and generative AI capabilities. Watsonx.ai gives the developer choice when selecting and customizing models as well as choice to deploy when, where and how they want. This software release includes several new features and foundation models as well as feature enhancements.
Below are the highlighted updates which will help accelerate the value of AI across your organizations. Full details of this release can be found in our “What's New” overview page.
Highlights of watsonx.ai v2.3
Watsonx.ai Model Gateway : Access Any Model, Anywhere
The Model Gateway allows clients to securely access and manage AI models from multiple providers in a rapidly evolving AI landscape. Currently this feature is available via the API only. This feature helps eliminate vendor lock-in by enabling seamless integration of internal models, curated models from the watsonx.ai Deploy on Demand catalog, and external models hosted by third parties. Acting as a single control plane, the gateway ensures governance and security through centralized access control, rate limiting, load balancing, and credential management in secure vaults. Its standardized OpenAI-compatible API simplifies adoption across diverse environments, whether on-premises or in the cloud. With easy onboarding via the SaaS Resource Hub and support for providers like Azure OpenAI, organizations can confidently leverage the latest AI innovations while maintaining compliance, cost efficiency, and operational resilience. Learn more in documentation
Generate Unstructured Synthetic Data
The Synthetic Data Generator now enables users to generate large, high-quality unstructured text datasets for foundation model tuning and evaluation - now supported both programmatically using the REST API as well as user interface to create jobs for generating unstructured synthetic data. With the Synthetic Data Generator, users can now create large, high-quality unstructured text datasets tailored to specific needs as well as use the unstructured synthetic datasets that are generated to tune and evaluate foundation models for specific use cases. Learn more in documentation
New Text Classification API
The new text classification method in the watsonx.ai REST API enables users to classify the document before extracting the textual content to use in a RAG solution. Users can classify their document with the classification API into one of several supported common document types without running a longer extraction task. By pre-processing the document, users can then customize the text extraction request to efficiently extract relevant details from the documents. Learn more in documentation
New Document Storage type for Text Extraction
Use the text extraction API with documents stored in the following data stores:
Learn more in documentation
Leverage New Foundation Models
We are bringing in a set of new foundation models that are now available to provide choice and flexibility for your generative AI use cases
See the full list of supported foundation models here
Prompt Multiple Vector Indexes in Prompt Lab
Leverage already created multiple vector index assets to chat seamlessly in a single conversation. This enables unified retrieval from all indexed data sources, delivering more comprehensive and context-rich answers without switching between assets. Perfect for RAG workflows, it ensures faster insights and a smoother experience when working with diverse datasets. Learn more in documentation
New Data Connections for AutoAI for RAG Experiments
New data connections for document collections and test data in AutoAI for RAG experiments. Supports the below additional data connections
-
Google Cloud Storage
-
Box
-
Dropbox
Enhancements for AutoAI for RAG experiments
Leverage the following features for AutoAI for RAG experiments:
-
Semantic chunking in AutoAI for RAG experiments
-
Chat API models in AutoAI for RAG experiments
-
Auto-deploy top pattern in AutoAI for RAG experiments
-
Multiple vector indexes in AutoAI for RAG experiments
-
SQL database schemas in AutoAI for RAG experiments
Learn more in documentation
Enhancements for Custom Foundation Models
Leverage from the following features for Custom Foundation Models
Prompt Tuning removal
The Prompt Tuning method is now removed which means users can no longer use prompt tuning as a method to tune foundation models. The deprecation was announced in version 2.2. Existing Prompt Tuning deployments will be removed on upgrading to v2.3 when. For details about alternative tuning methods, see <insert link for Planning for foundation model tuning>. Learn more in documentation
Runtime Environment Change : Migrating Away from Anaconda Packages
Starting with Runtime 25.1 (release expected by end of Q2 2026), there will be change in the way virtual environments are handled and packages are managed. Instead of a combination of conda virtual environments and conda package manager, there will now be a combination of Python virtual environments and pip package manager. Learn more about this change and migration steps in here
Sizing Estimate Change
Sizing estimate for watsonx.ai v2.3 has been updated to reflect recent architectural enhancements and feature additions. Important aspects to note are
-
Existing customers can still upgrade to v2.3 without adding VPCs if they stay within their current cluster capacity. i.e., if they are already being charged based on the cluster capacity which is less than the total of all VPCs of the services
-
The compute (vCPUs) for v2.3 is lesser than v2.2, so customers can install and use watsonx.ai on v2.3 with their existing capacity, but since the CPU limits have gone up, the entitlement (VPC) has increased
-
Scaling to new maximum limits may require additional VPCs, but base operation does not.
Learn more about this change in here