# Kimi K3 on Snowflake Cortex AI > Source: > Published: 2026-09-24 17:02:27+00:00 Kimi K3, Moonshot AI's open-weight model, is now available in private preview on Snowflake Cortex AI. Adding K3 to Cortex AI gives teams another option for matching the right model to the job, whether that means frontier reasoning for complex tasks or reserving higher-cost proprietary models for where they make the biggest difference. At launch, preview customers can use Kimi K3 through Snowflake Cortex AI Functions and Snowflake Cortex Inference. Support for Snowflake CoCo, Snowflake Cortex Agents, and Snowflake CoWork is coming soon. In [Moonshot's published evaluations](https://www.kimi.ai/blog/kimi-k3), K3 autonomously built a working GPU compiler from scratch, reproduced complex astrophysics research in two hours, and sustained multi-day coding sessions across large repositories. These results point to a model designed for work that takes many steps, uses more than one kind of input, and needs to carry context forward. The sections below cover what that looks like on Snowflake. ## Kimi K3: designed for long-running work Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model that activates 16 of 896 experts per token using Moonshot's Stable LatentMoE framework. Moonshot reports a 2.5x improvement in scaling efficiency over its predecessor, K2. In practice, that architecture targets work that takes several steps and uses more than one kind of input. Moonshot highlights long-running coding sessions that navigate large repositories, orchestrate terminal tools, and use screenshots to refine the result. Its research examples combine literature review, executable code, validation, and interactive visualizations. For teams, that maps to repository-level development, research synthesis, and analytical workflows where the model needs to carry context forward across many turns. These examples describe model-level capabilities. The inputs, tools, context limits, and output handling available to you depend on the Snowflake surface and configuration you use. Sources: Moonshot AI, [Kimi K3: Open Frontier Intelligence](https://www.kimi.ai/blog/kimi-k3) and [Hugging Face](https://huggingface.co/moonshotai/Kimi-K3%20) ## What you can do with Kimi K3 on Snowflake Kimi K3 adds an open-weight model option to Snowflake's model selection. During private preview, teams can evaluate it against their own tasks and quality requirements without setting up a separate inference deployment. **Coming soon**, support for Kimi K3 in Snowflake CoCo, Snowflake CoWork, and Cortex Agents will extend that evaluation to coding, research, and data agent workflows: - **Develop with Snowflake CoCo:** Repository-level work depends on more than the file in front of you. A developer may need to trace a call through several modules, understand the existing tests, and carry a change through to a working result. Kimi K3's focus on long-running coding makes this a useful area to evaluate in CoCo. Start a fresh session with Kimi K3 and give it a bounded task with explicit acceptance criteria. Moonshot recommends against switching an ongoing session from another model to K3 because the model is sensitive to how conversation history is preserved. - **Bring research into data agents:** With support in Snowflake CoWork and Cortex Agents, teams will be able to evaluate research and analysis workflows that draw on the data and tools available to their agent. For example, use Kimi K3 to compare customer research notes, identify conflicting requirements, and prepare a brief with the evidence behind each recommendation. ### Benchmark results The following snapshot showcases results from Moonshot AI's published evaluation table including coding, agentic, reasoning, and visual benchmarks. For full methodology, harness details, and footnotes, see [Moonshot AI's Kimi K3 model card](https://www.kimi.ai/blog/kimi-k3). | Benchmark | Kimi K3 result | |---|---| | DeepSWE | 67.5 | | ProgramBench | 77.8 | | Terminal-Bench 2.1 | 88.3 | | FrontierSWE | 81.2 | | BrowseComp | 91.2 | | DeepSearchQA | 95.0 | | ResearchRubrics | 76.2 | | GDPval-AA v2 | 1686 | | GPQA Diamond | 93.5 | | AA-LCR | 74.7 | | OmniDocBench | 91.1 | | PerceptionBench | 58.5 | Source: Moonshot AI ### Summarize research with Cortex AI Functions Cortex AI Functions bring model calls into SQL. An analyst can use AI_COMPLETE to compare research notes or summarize document text without writing a separate API client. This self-contained example uses fictional product interviews: ``` SELECT AI_COMPLETE( 'kimi-k3', 'Compare these fictional product research notes. Interview A: export jobs time out on large reports. Interview B: users need scheduled exports. Interview C: report definitions are hard to reuse. Group the needs, explain how they differ, and propose follow-up questions. Use only the supplied notes.' ) AS research_summary; ``` ### Call Kimi K3 with the OpenAI SDK Kimi K3 supports the Chat Completions request format, so developers can use the OpenAI SDK to call it through Snowflake's Cortex REST API. Set SNOWFLAKE_PAT to your Snowflake programmatic access token and SNOWFLAKE_ACCOUNT_URL to your account's HTTPS URL. ``` python import os from openai import OpenAI account_url = os.environ["SNOWFLAKE_ACCOUNT_URL"].rstrip("/") client = OpenAI( api_key=os.environ["SNOWFLAKE_PAT"], base_url=f"{account_url}/api/v2/cortex/v1", ) response = client.chat.completions.create( model="kimi-k3", messages=[{ "role": "user", "content": ( "Compare these fictional product research notes. " "Interview A: export jobs time out on large reports. " "Interview B: users need scheduled exports. " "Interview C: report definitions are hard to reuse. " "Group the needs, explain how they differ, and propose " "follow-up questions. Use only the supplied notes." ), }], ) print(response.choices[0].message.content) ``` ## What to know before you start - Conversation history matters. K3 was trained with preserved thinking history. Switching to K3 mid-session from another model, or using a harness that drops thinking content, can degrade output quality. Start fresh sessions when evaluating K3. - The model can be proactive. When K3 encounters minor issues or ambiguous intent, it may act on your behalf rather than asking for clarification. Use explicit constraints in system prompts and ensure you use appropriate agent approval permissions for deployments that require strict boundaries. - Moonshot states that K3's overall user experience "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol." Treat K3 as a strong open-weight option to evaluate, not a drop-in replacement for every workload. ## Get started The best model for a task depends on the task. In private preview, Kimi K3 gives teams an advanced open-weight model they can evaluate through Cortex AI Functions and Cortex Inference without provisioning separate infrastructure. Use model ID `kimi-k3`, run a task you already understand, and compare. Open-weight models at this scale used to require dedicated infrastructure and a team to run it. Now they're a model ID swap. Every new model on Snowflake's managed AI platform widens what teams can evaluate, and K3 adds one designed for work that takes many steps to finish. ## Learn more - **Build agentic applications with Cortex Agents:** Get started with[Cortex Agents](https://www.snowflake.com/en/developers/guides/getting-started-with-cortex-agents/) and learn how to build AI experiences grounded in your enterprise data. - **Get started with Snowflake CoWork:** Learn more about[Snowflake CoWork](https://www.snowflake.com/en/product/snowflake-cowork/) and try it out with our[getting started guide](https://www.snowflake.com/en/developers/guides/getting-started-with-cowork/) . - **Start building with Snowflake CoCo:** Learn more about[Snowflake CoCo](https://www.snowflake.com/en/product/features/cortex-code/) and get hands-on with the[CoCo Desktop getting started guide](https://www.snowflake.com/en/developers/guides/getting-started-with-coco-desktop/) . The Cortex REST API gives you access to leading frontier models through your preferred endpoint or SDK. All inference runs within the Snowflake perimeter, so your data remains secure and within your governance boundary. Explore [Cortex REST API](https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-rest-api) to learn more. **Forward-looking statements** This article contains forward-looking statements, including about our future product offerings, and are not commitments to deliver any product offerings. Actual results and offerings may differ and are subject to known and unknown risk and uncertainties. See our latest 10-Q for more information.