Kimi K3 on Snowflake Cortex AI

Snowflake Cortex AI now includes Moonshot AI’s Kimi K3, an open-weight model available in private preview. Teams can evaluate K3 through Cortex AI Functions and Cortex Inference to match tasks with appropriate models, reserving proprietary options for critical use cases. Support for additional Snowflake services like CoCo, Cortex Agents, and CoWork is planned for the future. The integration allows users to test K3’s capabilities directly within Snowflake’s managed platform.

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model using Moonshot’s Stable LatentMoE framework. Moonshot reports a 2.5x scaling efficiency improvement over its predecessor, K2, enabling sustained multi-step tasks such as long-running coding sessions and research synthesis. The model’s architecture targets workflows requiring multiple inputs, context retention, and iterative refinement, aligning with repository-level development and analytical tasks.

During private preview, teams can access K3 via Cortex AI Functions and Cortex Inference without provisioning separate infrastructure. Developers can call K3 using the OpenAI SDK through Snowflake’s Cortex REST API by setting specific environment variables. The model supports the Chat Completions request format, allowing analysts to integrate AI tasks into SQL workflows, such as summarizing documents or comparing research notes.

Cortex AI Functions enable model calls directly within SQL, while the Cortex REST API provides access to frontier models through preferred endpoints or SDKs. All inference runs occur within Snowflake’s perimeter, ensuring data remains secure and governed. K3’s addition expands the range of models teams can evaluate on Snowflake’s managed AI platform, offering an advanced open-weight option for complex, multi-step workflows.