Snowflake is expanding its Cortex platform to enable secure, governed custom model training, allowing enterprises to fine-tune AI models on proprietary data without moving information to external clouds.

Key Points

• Snowflake introduced Cortex Training to allow businesses to fine-tune open-weight models like Qwen and Mistral within a secure, governed environment.
• The platform utilizes multi-tenant GPU technology to increase training efficiency, potentially doubling the number of training runs per GPU budget.
• New engineering innovations, such as the ZoRRo zero-redundancy rollout, are being implemented to optimize reinforcement learning training workloads.
• Resolve AI is leveraging custom models to improve accuracy and performance for specialized tasks like autonomous site reliability engineering and software debugging.
• The strategy emphasizes combining general-purpose frontier models with domain-specific models to meet unique enterprise requirements for accuracy and latency.

Why it Matters

This shift allows organizations to deploy production-grade AI while maintaining strict data security and reducing the complexity of managing distributed GPU infrastructure. By enabling domain-specific model training, companies can achieve higher accuracy for specialized business use cases that general-purpose models often struggle to handle.