Snowflake Cortex just made custom LLMs a first-class citizen

Snowflake just shipped serverless fine-tuning for Cortex AI, its fully managed service for building AI applications. This moves enterprise ML development from a bespoke, infrastructure-heavy task to a managed workflow, making custom models accessible to any data engineer without needing to manage complex infrastructure.

what changed

The core update is the introduction of Cortex Fine-Tuning, a managed service that lets you customize popular large language models using your own data, all within the Snowflake security perimeter. This is a significant move because it brings the model to the data, not the other way around. Traditionally, using proprietary data for fine-tuning meant exfiltrating it to a separate environment, creating governance and security headaches. With this release, the training data is read directly from Snowflake tables, so it never leaves the platform's security boundary.

The service uses parameter-efficient fine-tuning (PEFT) to create customized adapters for pre-trained models. This approach avoids the high cost of training a large model from scratch while delivering better performance on specialized tasks than prompt engineering alone can achieve. The initial release supports models from providers like Meta and Mistral AI.

Developers can initiate and manage fine-tuning jobs through a simple SQL function or a no-code UI in the AI & ML Studio.

a practical workflow

Getting a custom model trained and running is now integrated directly into standard data workflows. You can launch a fine-tuning job with a single SQL command, pointing it to your training data stored in a Snowflake table.

The process is straightforward:

1. Prepare your data: Your training data must be in a Snowflake table with columns specifically named prompt and completion.
2. Launch the job: Execute the SNOWFLAKE.CORTEX.FINETUNE function, specifying a name for your new model, the base model you want to customize (e.g., 'mistral-7b'), and a SQL query to select your training data.
3. Monitor and deploy: The function returns a job ID that you can use to track the status. Once complete, the fine-tuned model is registered and immediately available for inference.
4. Run inference: Call your new model using the standard COMPLETE function, just as you would with any other pre-trained model in Cortex AI.

Here is what the SQL call looks like:

SELECT SNOWFLAKE.CORTEX.FINETUNE(
  'CREATE',
  'my_tuned_support_bot',
  'mistral-7b',
  'SELECT prompt, completion FROM support_tickets_train_set',
  'SELECT prompt, completion FROM support_tickets_validation_set'
);

This single command abstracts away all the underlying GPU provisioning, software environment configuration, and model management. The fine-tuned model's access is managed through standard Snowflake role-based access control (RBAC).

why it matters

This update signals a shift in how custom AI is built in the enterprise. By embedding fine-tuning directly into the data cloud, Snowflake is lowering the barrier to entry for creating domain-specific models. Data teams can now improve model accuracy for specialized tasks, like categorizing support tickets or summarizing financial documents, without a dedicated MLOps team managing GPU clusters.

This approach offers a compelling trade-off: you get the accuracy of a larger model from a fine-tuned smaller model, but with lower inference latency and reduced costs for repeated, specialized tasks. It makes building a custom LLM feel less like a research project and more like creating a materialized view.

Of course, this is a managed service, so you trade granular control for convenience. You won't be debugging CUDA drivers, but you are dependent on the models and architectures Snowflake chooses to support. Understanding the cost model is also key, as billing is based on the number of tokens processed during training and inference.

For builders working inside large enterprises, this is a significant development. The ability to securely and easily fine-tune models on your own data, without moving it, solves a major operational and security challenge. It makes custom AI a practical tool for any team that already works with data in Snowflake.

Sources

https://www.snowflake.com/
https://docs.snowflake.com/en/user-guide/cortex/cortex-fine-tuning