K2.8 Preview Operational

Kimi K2.8 Preview is Moonshot AI's primary model fully launched across Kimi Code and Kimi Work. The company says its overall performance is close to the flagship Kimi K3, with improved coding and Agent capabilities and significantly more efficient reasoning than K2.7. The model supports three reasoning levels—low, high, and max—as well as image and video input. Its 1M extended context window is available to all members.

Code generation and completion: Designed as the primary model for everyday development, optimized for well-defined tasks such as code completion and modification.
Agent capabilities: Coding and agent capabilities are comprehensively improved over K2.7 Code, enabling autonomous progress on multistep engineering tasks.
Three reasoning levels: Supports low / high / max reasoning effort, aligned with K3's reasoning levels, with max as the default.
Multimodal input: Supports image and video input for processing visual development assets.
1M extended context: All membership tiers have access to a context window of up to 1 million tokens, allowing large codebases to be processed at once.
Seamless upgrade: The Model ID remains kimi-for-coding, so no configuration changes are needed in clients or third-party tools.
Smart routing: When thinking is disabled for the K3 series, requests are automatically routed to the non-thinking version of K2.8.

Tiered reasoning control: The reasoning effort mechanism makes trade-offs in the reasoning budget—low sacrifices reasoning depth for speed, while max uses more reasoning tokens to improve success rates on complex tasks. At its core, this explicitly schedules the computing resource of “the number of reasoning tokens.”
Long-context processing: The 1M window relies on efficient attention implementation. K3's KDA hybrid linear attention and Attention Residuals approach, which explicitly reduces the computational cost of attention over long sequences, serves as an important technical reference for K2.8.
Request routing layer: On the product side, the scheduling system seamlessly switches session traffic between K3 and K2.8 based on model availability and the thinking toggle state, ensuring a continuous experience through the same entry point.

No configuration required: K2.8 Preview is fully launched, with kimi-for-coding upgraded seamlessly. Existing clients and third-party tools, including Claude Code, OpenCode, and Codex, require no configuration changes.
Available with any membership: Every tier from Adagio (free) through Allegro can use it, with direct access to the 1M context window.
Kimi Code CLI: After installing the CLI, sign in and start a session normally to use K2.8. Enter /model to view or switch the current model.
Adjust the reasoning level: Enter /effort to switch among low / high / max. Use a lower level to save tokens on simple tasks and max to improve success rates on complex tasks.
Third-party tool integration: Set the tool's model ID to kimi-for-coding. Allegretto members and above who prioritize speed can switch to kimi-for-coding-highspeed.
Kimi Work: Start a task directly in Kimi Work, and the platform will automatically route it to K2.8 Preview.

Near-flagship performance: Overall performance is close to K3, while coding and Agent capabilities are comprehensively improved over K2.7 Code and reasoning is significantly more efficient.
Zero-barrier access: Every membership tier, including the free Adagio tier, can use it, overcoming K3's starting price of ¥99.
Faster responses: User testing reports that it is noticeably faster than K3, making it suitable for high-frequency interactive scenarios such as code completion.
More predictable behavior: Compared with K3's tendency to be overly proactive on ambiguous tasks and its sensitivity to previous thinking, K2.8 is positioned for everyday development with more stable behavioral boundaries.
Seamless migration: The Model ID remains kimi-for-coding, allowing clients and third-party tools to upgrade without changes.
Flexible reasoning levels: Freely switch among low / high / max to save tokens on simple tasks and preserve quality on complex ones.
Multimodal capabilities: Supports image and video input, while the 1M window can take in an entire large codebase at once.

Understanding and refactoring large codebases: The 1M context can take in an entire repository at once for cross-file dependency analysis, batch refactoring, and architecture-level modification suggestions.
Everyday code completion and development: Positioned as an “everyday workhorse,” it is more predictable and responsive than K3, making it suitable for frequent completion, debugging, and small feature iterations.
Multimodal development assistance: Supports image and video input, allowing users to directly provide screenshots of error interfaces, bug recordings, or design references for generating frontend code.
Agent-powered automation workflows: Coding and Agent capabilities are comprehensively improved over K2.7 and can support multistep task orchestration, including the complete workflow of running tests, fixing failed cases, and submitting a PR.
Long-document and log analysis: The 1M window can accommodate extensive logs, technical documents, or meeting records for full-text search-style question answering and summary extraction, with no additional cost for members.