Kimi AI ✓

Stronger long-horizon coding sustains extended engineering tasks, navigates massive repositories, and coordinates terminal tools with minimal supervision.

Native visual reasoning combines screenshots, images, and video with coding for frontend development, game creation, CAD, and iterative visual refinement.

A 1-million-token context window enables much larger codebases, documents, and persistent agent workflows within a single working context.

More capable knowledge-work agents handle research, analysis, spreadsheets, reports, visualizations, and other production-oriented workflows end to end.

New KDA, Attention Residuals, and highly sparse MoE architecture deliver roughly 2.5x better scaling efficiency than K2 while reaching 2.8 trillion parameters.

Stronger coding turns simple prompts into polished interfaces, full-stack websites, authentication flows, interactions, and database-backed applications.

More reliable long-horizon execution stays aligned across complex multi-step tasks, makes fewer unnecessary changes, and corrects mistakes during execution.

Upgraded agent coordination lets specialized agents work in parallel across research, analysis, writing, websites, documents, slides, and spreadsheets.

Reusable skills can be extracted from high-quality documents and applied to future tasks for more consistent, scalable outputs.

Improved presentation generation produces editable, professional decks from prompts, screenshots, templates, and mixed-format source material.

Major improvements in agentic intelligence, enabling more autonomous task execution.

Further enhancements to reasoning depth, planning, and multi-step problem solving.

Optimized performance for complex workflows and large-scale context handling.

Added multimodal capabilities, enabling understanding of text, images, and code.

Significantly improved reasoning, mathematics, and coding performance.

Enhanced long-context comprehension and response consistency.