MoonshotAI: Kimi K2 0905
Model Type
Proprietary Model
API access only
Recommended Use Cases
Try Kimi K2 0905
Kimi K2 Instruct 0905 is the September 2025 update to Kimi K2, featuring doubled context length (256K tokens), enhanced agentic coding intelligence, and improved frontend development capabilities.
Overview
Kimi K2 0905 builds on the original Kimi K2 with significant improvements in coding performance, context handling, and agentic capabilities. The update doubles the context window from 128K to 256K tokens and introduces MuonClip optimizer improvements for more stable training and better generalization across coding scaffolds.
Key Improvements Over K2
- Doubled context window: Extended from 128K to 256K tokens
- Enhanced agentic coding: Better performance on real-world coding agent tasks
- Improved frontend development: More aesthetic and practical web, 3D, and UI outputs
- Better scaffold generalization: Consistent performance across different agent frameworks
Technical Specifications
| Specification | Value |
|---|---|
| Total Parameters | 1T |
| Active Parameters | 32B |
| Architecture | MoE (Mixture-of-Experts) |
| Context Length | 256K tokens |
| Format | block-fp8 |
Capabilities
- Full-stack development with polished design aesthetics
- Multi-file codebase navigation and refactoring
- Async logic and complex architectural patterns
- Bug fixing in large repositories (tested on 38K+ line Rust codebases)
- React project refactoring and modernization
Tool Calling
K2 0905 has strong tool-calling capabilities with support for:
- Function schema definition
- Autonomous tool selection
- Multi-turn tool interactions
- OpenAI and Anthropic-compatible API formats
Availability
- API: Moonshot AI platform (platform.moonshot.ai)
- Open Weights: Hugging Face (moonshotai/Kimi-K2-Instruct-0905)
- Third-party providers: OpenRouter, Fireworks, NVIDIA NIM, Together AI
- Local deployment: vLLM, Ollama
Role in Series
Moonshot AI's Kimi models offer different capabilities:
- Kimi K2: Base model, 128K context
- Kimi K2 0905: Enhanced coding, 256K context (this model)
- Kimi K2 Thinking: Extended reasoning with tool orchestration
- Kimi K2.5: Maximum capability with vision and Agent Swarm
Kimi K2 0905 replaces the original Kimi K2 for most coding and agentic use cases.