Moonshot AI’s Kimi K3: A 2.8T Parameter Beast for Coding & Agentic Work

Moonshot AI has raised the bar again with Kimi K3, a massive 2.8-trillion-parameter sparse Mixture-of-Experts (MoE) model that’s making waves as a powerful open-weight frontier contender. Designed for long-horizon coding, deep reasoning, and knowledge work, K3 features a groundbreaking 1-million-token context window, native vision capabilities, and innovative architecture improvements like Kimi Delta Attention and Attention Residuals for better efficiency.

In practical testing, Kimi K3 shines at agentic tasks — redesigning landing pages, analyzing codebases, and building full features with strong instruction following. It excels particularly in frontend development (topping Arena.ai’s Frontend Code Arena) and sustained engineering sessions, making it ideal for complex projects that require maintaining context across thousands of files or extended workflows.

Standout Features:

• Massive Scale — 2.8T parameters with smart expert activation
• Ultra-Long Context — 1M tokens for repository-scale understanding
• Multimodal — Native image and video understanding
• Agentic Strength — Excellent at planning, tool use, and iterative development

While still new (with full open weights expected soon), Kimi K3 delivers competitive performance against top closed models at more accessible pricing, positioning it as a strong choice for developers and teams focused on coding agents and production-grade AI applications.