Snowflake Adds Kimi K3 to Cortex AI in Private Preview
Teams can test Moonshot AI's open-weight model through Cortex AI Functions and Inference, while agent and coding surfaces remain future work.
Edited by Tyronne Panaino
Snowflake announced on September 24 that Moonshot AI's Kimi K3 is now available in private preview through Snowflake Cortex AI. Preview customers can use the open-weight model through Cortex AI Functions and Cortex Inference, giving enterprise teams a managed place to test it without first standing up a separate inference deployment. The change matters to Snowflake users evaluating long-running research or development work, but it is not a general-availability release and several higher-level product integrations are not yet live.
What is available now
The Snowflake announcement draws a clear boundary around the present release. Kimi K3 is available to private-preview customers in Cortex AI Functions and Cortex Inference. Snowflake says those customers can compare the model against their own tasks and quality requirements without provisioning a separate inference system.
That makes the current release an evaluation path inside Snowflake's managed AI platform, not evidence that every Snowflake account can select Kimi K3. Private preview normally implies controlled access, and the announcement does not provide a general-availability date or a public eligibility process. Teams should confirm access in their own account before planning production work around the model.
Agent and coding integrations are still pending
Snowflake separately lists support for Cortex Agents, Snowflake CoCo and Snowflake CoWork as coming soon. Those surfaces are important because they would place Kimi K3 inside agent, coding and research workflows rather than limiting it to the two inference interfaces available in the preview. The announcement does not commit to dates for those additions and includes an explicit forward-looking-statement notice that future product offerings may change.
This split is the practical delta to watch: model calls are available now in the named preview interfaces, while the surrounding agent products remain prospective. An evaluation completed through Cortex AI Functions or Cortex Inference should not be treated as proof that the same behavior, tools or context handling will carry over unchanged when the other products arrive.
Session history and approval controls need testing
Snowflake passes on two operational cautions for evaluators. First, Moonshot recommends starting a fresh Kimi K3 session instead of switching an existing conversation from another model, because the model is sensitive to how conversation history is preserved. That makes session construction part of the test plan rather than an invisible implementation detail.
Second, the announcement says Kimi K3 can respond proactively when it meets minor problems or ambiguous intent. Snowflake advises explicit system constraints and suitable approval permissions for deployments that require strict boundaries. Teams evaluating the preview should therefore test not only answer quality, but also whether the model pauses, escalates or acts in the way their workflow expects.
Those statements are vendor guidance, not independent evidence that a given configuration is safe or reliable. The announcement does not provide tenant-level security testing, production incident data or independent comparisons of Snowflake-hosted Kimi K3. It also does not establish that preview behavior will remain identical at general availability.
What to verify next
The next verifiable checkpoints are broader tenant access, dated availability for CoCo, CoWork and Cortex Agents, and documentation that fixes the precise model, tool and context behavior exposed by each surface. Customers should also compare Kimi K3 against tasks they already understand, using acceptance criteria and approval boundaries that can reveal failures rather than relying on model-launch claims.
Status
Confirmed. Snowflake's official product post establishes the private preview and the two currently available Cortex AI surfaces. Internal confidence is medium because the evidence is first-party, the agent integrations are forward-looking, and no independent Snowflake tenant test was available in this run.
Sources
Update note: Last reviewed 2026-09-25. We will revise this post if Snowflake changes preview access, publishes dates for the pending integrations or documents general availability.
Sources
- Snowflake — Kimi K3 on Snowflake Cortex AI — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.