Evergreen Tools
|
← Back to AI Tools

Kimi K2.8 Preview Multimodal Coding Model

A preview model that went live in Kimi Code on 11 September 2026, bringing vision and million-token context to the Kimi family for the first time

Model Overview

Specs, access and pricing details below

Features

  • ✓ Brings vision and multimodal input to the Kimi family for the first time, handling text, images and video for document analysis and visual question answering
  • ✓ Every Kimi Code membership tier gets up to a 1M-token context window, up from 262,144 tokens on K2.7 Code
  • ✓ Thinking levels of low / high / max (max by default), matching Kimi K3
  • ✓ Keeps the same API model ID (kimi-for-coding), so existing users and third-party tools need no configuration changes
  • ✓ Sits between the coding-focused Kimi K2.7 Code and the flagship Kimi K3 with coding and agent ability said to approach K3; proprietary and API-only

How to Use

  1. Open Kimi Code on the web or the Kimi Code CLI
  2. Confirm K2.8 Preview has rolled out to your preview channel (the model ID is unchanged, so no reconfiguration is needed)
  3. Upload or paste text, images or video and pick a thinking level (max by default)
  4. Compare coding or agent tasks against K2.7 Code for quality and token usage

FAQ

What is Kimi K2.8 Preview?

A preview model Moonshot AI rolled out in Kimi Code on 11 September 2026, positioned between the coding-focused K2.7 Code and the flagship K3, and the first to bring vision and multimodal input to the Kimi family.

What inputs does it support?

Text, images and video, for use cases such as document analysis and visual question answering; the earlier K2.7 series was text-only.

How large is the context window?

Every membership tier gets up to a 1M-token context window, up notably from the 262,144 tokens on K2.7 Code.

Do I need to change my configuration?

No. K2.8 Preview keeps the existing API model ID (kimi-for-coding), so current users and third-party tools can use it without changing any settings.

How do the thinking levels work?

There are three levels, low / high / max, matching K3, with max as the default; with thinking turned off, requests to the K3 series and K2.8 Preview are served by K2.8 Preview without thinking.

Is it open source?

No. K2.8 Preview is a proprietary model available only through the API and Kimi's own products, with no downloadable open weights.