Kimi K2-Instruct
Active Parameters
1T
Context Length
128K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
Modified MIT License
Release Date
11 Jul 2025
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
140x RTX 4090
24GB VRAM
Datacenter
34x NVIDIA A100
80GB VRAM
Apple Silicon
30x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
162x RTX 4090
24GB VRAM
Datacenter
39x NVIDIA A100
80GB VRAM
Apple Silicon
35x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
Rank
#43
| Benchmark | Score | Rank |
|---|---|---|
QA Assistant | 0.975 | 🥉 3 |
General Knowledge | 0.895 | ⭐ 4 |
Summarization | 0.93 | 6 |
Coding | 0.59 | 16 |
Graduate-Level QA | 0.751 | 36 |
Professional Knowledge | 0.79 | 40 |
Rankings
Overall Rank
#43
Coding Rank
#44
About Kimi K2-Instruct
Kimi K2-Instruct is an advanced Mixture-of-Experts (MoE) language model developed by Moonshot AI. This model incorporates 1 trillion total parameters, with approximately 32 billion parameters activated during each inference pass. Its core purpose is to deliver state-of-the-art agentic intelligence, facilitating sophisticated tool utilization, advanced code generation, and autonomous problem-solving across various domains. As a post-trained instruction-following variant, Kimi K2-Instruct is optimized for general-purpose conversational tasks and complex agentic workflows, operating as a reflex-grade model designed for direct application.
The architectural design of Kimi K2-Instruct features a Mixture-of-Experts paradigm, leveraging 384 specialized experts, with 8 active experts dynamically selected per token during inference. The model comprises 61 layers and employs a Multi-head Local Attention (MLA) mechanism with 64 attention heads. A key innovation in its training methodology is the MuonClip optimizer, developed by Moonshot AI, which ensures training stability at the expansive scale of 15.5 trillion tokens. The architecture prioritizes long-context efficiency, supporting a substantial context window of 128,000 tokens. The activation function employed within the model is SwiGLU, complemented by Rotary Position Embeddings (RoPE).
Kimi K2-Instruct is engineered for demanding applications, including complex, multi-step reasoning tasks and analytical workflows that necessitate profound comprehension. Its capabilities encompass advanced code generation, ranging from foundational scripting to intricate software development and debugging, along with robust support for multilingual applications. The model exhibits strong tool-calling capabilities, enabling it to autonomously interpret user intentions and orchestrate external tools and APIs to accomplish intricate objectives. Practical use cases include automating development workflows, generating comprehensive data analysis reports, and facilitating interactive task planning by seamlessly integrating multiple external services.
Technical Specifications
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
64
Key-Value Heads
64
Attention Head Dimension
-
Position Embedding
ROPE
RoPE Theta
50,000
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
7,168
Number of Layers
61
FFN Intermediate Size (Dense)
2,048
Multi-Token Prediction Heads
0
Tokenizer
Vocabulary Size
163,840
Mixture of Experts
Total Expert Parameters
32.0B
Number of Experts
384
Active Experts
8
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
1
About Kimi K2
Moonshot AI's Kimi K2 is a Mixture-of-Experts model featuring one trillion total parameters, activating 32 billion per token. Designed for agentic intelligence, it utilizes a sparse architecture with 384 experts and the MuonClip optimizer for training stability, supporting a 128K token context window.