Kimi
Moonshot AI
Moonshot AI's Kimi series is known for pioneering long-context understanding and efficient MoE architectures. K2 became a major open-source competitor in 2025.
Models
Kimi K2 Thinking
1T (32B active) parameters
Reasoning and tool-using thinking agent. Can execute 200-300 sequential tool calls without human interference. State-of-the-art agentic capabilities.
- 200-300 sequential tool calls
- LiveCodeBench-v6: 83.1%
- Advanced agentic reasoning
Benchmarks
humaneval
91%
gsm8k
96.5%
Kimi Linear
48B (3B active) parameters
Uses Kimi Delta Attention (KDA) for efficient long-context processing. Reduces memory usage and improves generation speed at longer context windows.
- 1M token context window
- Novel Delta Attention mechanism
- Efficient memory usage
Benchmarks
mmlu
78%
Kimi K2
1T (32B active) parameters
1 trillion parameter MoE with 32B active. State-of-the-art open-source performance on coding benchmarks. Trained for $4.6M, rivaling ChatGPT and Claude.
- SOTA open-source coding performance
- Trained for only $4.6M
- Beats GPT-4o on multiple benchmarks
Benchmarks
mmlu
87.5%
humaneval
90.5%
gsm8k
95%
Kimi-Dev
72B parameters
Coding-focused model based on Qwen2.5-72B. State-of-the-art among open source on SWE-bench Verified.
- SOTA open-source on SWE-bench Verified
- Built on Qwen2.5-72B foundation
- Specialized for software development
Benchmarks
humaneval
89%
Kimi-VL
16B (3B active) parameters
Open-source vision-language MoE model. Efficient multimodal understanding with only 3B active parameters.
- 16B MoE with 3B active
- Efficient multimodal processing
- Apache 2.0 license
Kimi K1.5
Unknown parameters
First major Kimi reasoning model. Matches OpenAI o1 in math, coding, and multimodal reasoning. Uses reinforcement learning for dynamic learning.
- Matches OpenAI o1 performance
- Free with no usage limits
- Long-CoT and short-CoT modes
Benchmarks
mmlu
85%
gsm8k
93%
math
85%