Since its release in July 2023, IBM watsonx.ai™ has been enabling businesses to train, validate, tune and deploy AI models. With its cutting-edge generative AI and machine learning (ML) features and capabilities, watsonx.ai is designed to help businesses capitalize on the opportunities of generative AI and foundation models, while allowing them to mitigate risks and drive trust, performance and flexibility of AI outcomes.

IBM’s approach to AI is based on four core pillars: open, trusted, targeted and empowering. Our AI and data platform, watsonx, offers builders control and portability, and is designed for the enterprise. The watsonx platform allows users to tune and train foundation models all from the same interface with end-to-end lifecycle governance and integrate enterprise applications and data across any cloud environment.

We’re excited to announce a new feature update to watsonx.ai that delivers an open framework, giving users access to a catalogue of built-in models and patterns that can be seamlessly extended through a “bring your own model” (BYOM) capability. With watsonx.ai 1.1.4 software release, you can do even more with our enterprise AI studio: upload and deploy your own custom foundation models.

There are many reasons to import a custom foundation model, all driven by the unique needs of your organization. Ultimately, it boils down to a specific foundation model that is optimal for the task at hand but currently resides outside of watsonx.ai. For instance, you may need support for a language that is not currently available in the watsonx.ai foundation model library. Or your organization may have invested resources to fine-tune a model to optimize it for your specific industry or business domain. The BYOM approach provides users greater flexibility in how you select and utilize the right model to meet your specific generative AI use cases and technical tasks.

In addition to working with foundation models that are curated by IBM, including open source, third party, and IBM-developed watsonx foundation models, you can now upload and deploy your own custom foundation model to accomplish a range of industry or domain-specific generative AI tasks. For instance, a common task for many clients is summarizing customer service transcripts or generating personalized outbound emails. Another popular use case is to tune a Large Language Model (LLM) for a specific language, or with specialized labeled data to customize a model to an industry or business domain. By deploying this custom model into watsonx.ai software, you can leverage it within your applications as well as have access to the platform’s enterprise-ready governance features. Further, with our on-premises solution, you’re bringing the model closer to where your data resides, mitigating your risk exposure.

In this release, watsonx.ai will initially support the base versions or customizations of foundation models for natural language and programming language generation within our supported model architecture types. By deploying a custom foundation model to watsonx.ai, you can work with a model that best fits your project and business needs. One place to find models is Hugging Face, a repository for open-source foundation models used by many model builders. Or you can bring in models that you’ve already fine-tuned from your own environment. Note that you cannot further tune the custom model once it’s deployed in watsonx.ai as part of this initial software release.

The supported model architecture types that you can import into watsonx.ai include the following:

• bloom
• codegen
• falcon
• gpt_bigcode
• gpt_neox
• gptj
• llama2
• mixtral
• mistral
• mt5
• mpt
• t5

It’s possible that a model you are familiar with may have a name that is more commonly known and not represented on the list above. For instance, the model called starcoder is based on the architecture type called gpt_bigcode. Therefore, if you don’t see a model of interest included on the list above, check the model’s information card to learn which architecture type it is based on before starting an import. Once the model is imported and deployed, prompt engineers and model users can interact with the custom model as they would with other models in the watsonx.ai studio, for instance:

• Using the Prompt Lab to build and test prompts, including creating reusable prompt templates
• Programmatically accessing the model using REST API calls

For more information about watsonx.ai, IBM’s next-generation enterprise studio for AI builders to train, validate, tune and deploy generative AI and ML models, see below.