Databricks Integrates Moonshot AI's Kimi K3

Databricks has announced the availability of Moonshot AI's Kimi K3, an advanced open-weight model, on its platform through the Foundation Model API. Initially hosted in the United States, the model is accessible with native support for AWS and GCP workspaces, alongside ADI access for Azure Databricks. This integration allows enterprise users to deploy Kimi K3 within their secure Databricks perimeter, fully governed by the Unity AI Gateway and backed by comprehensive zero data retention coverage.

The addition of Kimi K3 offers practitioners a powerful alternative to proprietary models like Anthropic Claude, OpenAI GPT, and Google Gemini. In performance evaluations, Kimi K3 secured the fourth position with a score of 57 in the Artificial Analysis Intelligence Index, placing it well above average against comparable models. Databricks' internal enterprise benchmarks also confirm that the model consistently performs at the level of leading proprietary systems, particularly in tasks involving coding, agentic reasoning, and document understanding.

For developers and platform teams, this release changes the economics of deploying large-scale AI. Kimi K3 delivers a 50% to 72% reduction in cost-per-task compared to similar proprietary models. This price drop makes previously expensive workloads, such as continuous coding agents and massive document processing pipelines, commercially viable. Through the Unity AI Gateway, administrators can manage these costs using spend controls to set budgets per user, team, workspace, or account, while tracking expenses via a Cost Analytics dashboard.

Practitioners can connect Kimi K3 directly to their governed enterprise data using Unity Catalog, allowing AI agents to query lakehouses and access feature stores. The model can be tested immediately in the Databricks AI Playground with zero setup, or integrated into complex workflows using Agent Bricks. Because Databricks hosts the model directly, all data remains within the customer's existing security perimeter, maintaining identical governance policies across both open-weight and proprietary models.