MoonshotAI: Kimi Dev 72B
Model Type
Open Weight Model
72B parameters
Recommended Use Cases
Text Generation
Kimi Dev 72B is Moonshot AI's open-source coding model released in June 2025, achieving state-of-the-art performance among open-source models on SWE-bench Verified with a 60.4% resolve rate.
Kimi-Dev-72B achieves a new state-of-the-art on SWE-bench Verified among open-source models.
- Moonshot AI
Overview
Kimi Dev 72B is a coding-specialized LLM built on Qwen2.5-72B and optimized through large-scale reinforcement learning. The model autonomously patches real repositories in Docker environments and only receives rewards when the entire test suite passes, ensuring robust and production-ready code completions.
Key Features
- 60.4% SWE-bench Verified: State-of-the-art among open-source coding models
- RL-based training: Rewards only for complete test suite passing
- Dual roles: Trained as both BugFixer and TestWriter
- Test-time self-play: Coordinates bug-fixing and test-writing for verification
- MIT License: Fully open source for commercial and personal use
Technical Specifications
| Specification | Value |
|---|---|
| Base Model | Qwen2.5-72B |
| Parameters | 72B |
| Context Length | 131K tokens |
| Mid-training Data | ~150B tokens |
| License | MIT |
Availability
- Open Weights: Hugging Face (moonshotai/Kimi-Dev-72B)
- GitHub: MoonshotAI/Kimi-Dev
- Inference: vLLM (8x GPU tensor parallel recommended)
- Third-party providers: OpenRouter, SiliconFlow
Role in Series
Kimi Dev 72B is a specialized coding model separate from the main Kimi K2 series:
- Kimi Dev 72B: Coding specialist based on Qwen2.5, SWE-bench focused (this model)
- Kimi K2 0905: General agentic model with strong coding, 1T parameters
- Kimi K2.5: Multimodal with coding from vision capabilities