Next upAI x Bio Pitch Contest
News

Snowflake adds Kimi K3 to Cortex AI in private preview

Snowflake has made Moonshot AI’s Kimi K3 available to explicitly enabled accounts through Cortex AI Functions and Cortex Inference, while integrations with its agent and coding products remain forthcoming.

D
Sep 26, 2026 · 2 min read

Snowflake has made Moonshot AI’s Kimi K3 model available in private preview through Cortex AI Functions and Cortex Inference. According to Snowflake, explicitly enabled customers can evaluate the open-weight model without setting up a separate inference deployment.

Access is more restricted than Snowflake’s public preview of Grok 4.6 on Cortex AI. Snowflake’s documentation says private-preview models are available only to accounts it explicitly enables and are not suitable for production workloads.

The preview does not cover every Snowflake AI product. Support for CoCo, CoWork and Cortex Agents is coming soon, Snowflake said, while its announcement cautions that future offerings are not commitments and may differ. The company did not give delivery dates for those integrations.

Customers can call Kimi K3 from SQL with AI_COMPLETE and the model identifier kimi-k3. Snowflake’s Cortex AI Functions availability table lists the model as a private preview.

Kimi K3 is also accessible through the Cortex REST API. Snowflake says the model supports the OpenAI Chat Completions request format, so customers can use the OpenAI SDK. The Cortex REST API documentation also lists kimi-k3 as a private-preview model. Snowflake says inference runs within its perimeter.

Moonshot describes Kimi K3 as a native multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated parameters and a 1,048,576-token context length. Moonshot says its Stable LatentMoE design improves overall scaling efficiency by about 2.5 times compared with Kimi K2. That figure and the benchmark results in Moonshot’s model card are vendor-reported rather than independently replicated findings.

Moonshot says Kimi K3 was trained with preserved thinking history and requires complete assistant messages, including reasoning content and tool calls, during multi-turn use. Snowflake advises customers to start a fresh session instead of switching an existing session from another model. The announcement does not disclose pricing or a general-availability date for Kimi K3.

More news