Snowflake iconSnowflakeSep 24, 2026 ~6 min source read

Kimi K3 arrives in private preview on Snowflake Cortex AI

Moonshot AI’s 2.8-trillion-parameter open-weight model Kimi K3 is available to Snowflake private preview customers through Cortex AI Functions and Cortex Inference, offering a model built for long-running coding, research, and multi-input workflows.

Kimi K3 on Snowflake Cortex AI

Share this story

Send the public story page.

Useful takeaways from this story.

Kimi K3 (2.8T parameters) is in private preview on Snowflake Cortex AI via Cortex AI Functions and Cortex Inference.

Architecture: mixture-of-experts (activates 16 of 896 experts per token) using Moonshot’s Stable LatentMoE and reported 2.5x scaling efficiency vs. K2.

Designed for long-running, multi-step work — repository-level coding, multi-turn research, and agent workflows — with upcoming support for Snowflake CoCo, CoWork, and Cortex Agents.

# What was announced

Snowflake announced that Kimi K3, Moonshot AI's open-weight large model, is now available in private preview on Snowflake Cortex AI. At launch preview users can call K3 through Cortex AI Functions and Cortex Inference. Snowflake says support for CoCo (code workflows), CoWork (collaborative research), and Cortex Agents is coming soon.

# What Kimi K3 is

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts (MoE) model. It activates 16 of 896 experts per token and uses Moonshot's Stable LatentMoE framework. Moonshot reports a 2.5x improvement in scaling efficiency compared with Kimi K2. The model is presented as suited to tasks that require carrying context across many turns, combining multiple input types, and executing multi-step reasoning.

# Why Snowflake customers might evaluate it

K3's design targets workflows that span files, sessions, or documents rather than single-turn prompts. Snowflake highlights use cases such as:

  • Repository-level development that traces calls across modules, runs tests, and maintains session state across a long coding task.
  • Research workflows that combine literature review, executable code, validations, and visualizations.
  • Agentic data workflows that coordinate tools and data across turns inside Snowflake's environment.

Private preview lets teams compare K3 to other models without a separate inference deployment, using their own tasks and quality requirements.

# How to call Kimi K3 on Snowflake

Snowflake shows multiple integration points:

  • SQL: Cortex AI Functions bring model calls into SQL. The announcement includes an AI_COMPLETE example that uses 'kimi-k3' to compare fictional product research notes inside a single SQL call.

# Practical considerations

  • Session state sensitivity: Moonshot recommends starting fresh sessions with K3 for bounded tasks and avoiding switching in the middle of an ongoing session with a different model, because K3 is sensitive to how conversation history is preserved.
  • Feature rollout: At announcement time K3 is available in private preview via Cortex AI Functions and Cortex Inference. Snowflake lists upcoming support for CoCo, CoWork, Cortex Agents, and CoWork integrations that will expand evaluation into coding and agent workflows.

# Short summary

More context around this story.

Introducing Kimi K3 on Amazon Bedrock
Amazon iconAmazonSep 18, 2026

Introducing Kimi K3 on Amazon Bedrock

Kimi K3 from Moonshot AI is now available on Amazon Bedrock, giving you a powerful new open-weight option for coding and knowledge work. It offers native vision, a 1-million-token context window, and explicit prompt caching to reduce latency and input costs.

OpenAI GPT-6 AstraがSnowflake Cortex AIで利用可能に
Snowflake iconSnowflakeSep 9, 2026

OpenAI GPT-6 AstraがSnowflake Cortex AIで利用可能に

OpenAIのGPT-6 AstraがSnowflake Cortex AIでプライベートプレビューになりました。Snowflake CoCo、CoWork、Cortex Agents、AI関数、Cortex推論全体で、その最先端の推論とコンピューター使用を活用できます。これらはすべてSnowflakeのガバナンス境界内で実行されます。

Loading more related stories...

Keep reading in the app

Open the app view to save this story, compare related coverage, and continue from the same source.

Open in app