Enable streamlined and validated import and deployment of custom foundation models on watsonx.ai SaaS

Who would benefit from it?

Straker plans to run Tiri, its own Gemma 3-based model, as a custom foundation model on watsonx.ai SaaS for SwiftBridge AI v2, Straker's IR translation solution. A streamlined and validated import and deployment flow would help Straker proceed with its planned production launch while reducing deployment effort and configuration risk.

This capability would also benefit partners and enterprise customers that bring their own models to watsonx.ai SaaS, especially teams that need a repeatable and validated deployment process across multiple models, environments, or model versions.

Why is it useful?

The current model upload and deployment process is entirely manual. There is no registry-based import that allows a user to point watsonx.ai to a Hugging Face or equivalent model reference.

The current working process requires users to:

• Download the model weights
• Restructure the files into the layout expected by the runtime
• Upload the files to a Cloud Object Storage bucket
• Attach the Cloud Object Storage bucket to the watsonx project as a connection. Standard bucket credentials are not sufficient, as the connection requires specific credential types.
  Import the model asset
• Create the deployment

Several steps have failure modes that can currently be identified only through trial and error:

• Job-type trap: Deployments default to text generation. Unless the deployment declares the chat function, the chat endpoint is not served and requests return a 404 response.
• Chat template required in two configurations: The model's chat template must be present in both the checkpoint's tokenizer configuration and the deployment's serving configuration. Otherwise, the template can silently fail to apply, resulting in degraded output that may be mistaken for a model-quality problem.
• Documentation lag: Checkpoint formats accepted by the current runtime are documented as unsupported, while the documented formats reflect an older runtime. Straker had to identify and validate the working combination through trial and error.

Custom foundation models can be deployed and served successfully, but the current process is entirely manual and error-prone. This process increases deployment effort and creates a risk of configuration errors that may not be detected before deployment. Silent failures are particularly problematic because they can produce degraded model output without clearly identifying the deployment configuration as the cause.

How should it work?

watsonx.ai SaaS should provide a streamlined and validated bring-your-own-model flow for custom foundation models, including:

• Registry-based import from Hugging Face or an equivalent model registry, and/or direct model upload
• Automated handling or clear guidance for the model file structure and Cloud Object Storage upload
• Validation and guidance for the Cloud Object Storage connection and the specific credential types required by watsonx.ai
• Pre-deployment validation of checkpoint format, chat template configuration, and endpoint functions
• Clear and actionable validation errors before deployment
• Documentation versioned against the runtime actually deployed

The flow should reduce the current manual steps and detect incompatible formats, missing configuration, or invalid connection settings before deployment, making custom model deployment repeatable and reliable.