Kimi K2-Base
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
#64
| Benchmark | Score | Rank |
|---|---|---|
Summarization | 0.93 | 6 |
General Knowledge | 0.878 | 7 |
Graduate-Level QA | 0.758 | 34 |
General Text | 1418 | 74 |
Rankings
Overall Rank
#64
Coding Rank
-
About Kimi K2-Base
Kimi K2-Base is a foundational large language model developed by Moonshot AI, designed for researchers and developers who require a customizable base for specific applications. It is engineered to facilitate agentic tasks, encompassing advanced code generation, multi-step problem-solving, and the autonomous utilization of external tools and APIs. This model provides a robust platform for developing tailored AI systems across diverse domains, such as legal analysis, scientific research, and specialized conversational interfaces.
Architecturally, Kimi K2-Base is a Mixture-of-Experts (MoE) transformer model. It comprises a total of 1 trillion parameters, with 32 billion parameters activated during each inference. The architecture integrates 384 specialized experts, with 8 experts dynamically selected per token to process inputs. A key innovation in its development is the MuonClip optimizer, proprietary to Moonshot AI, which addresses training instability in large-scale models by mitigating exploding attention logits. The model's internal structure includes 61 layers, an attention hidden dimension of 7168, and employs 64 attention heads along with SwiGLU activation functions.
The Kimi K2-Base model supports a substantial context window of 128,000 tokens, allowing it to process and analyze extended inputs and multi-turn interactions effectively. This design contributes to its efficiency in inference and makes it suitable for applications requiring extensive contextual understanding. Its optimization for agentic intelligence signifies its capability to interpret goals and execute complex workflows without continuous human intervention. The model was pre-trained on an extensive dataset of 15.5 trillion tokens, supporting its performance across various knowledge, reasoning, and coding tasks.
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.