Active Parameters
1T
Context Length
128K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
69 KB/tok
License
Modified MIT License
Release Date
11 Jul 2025
Knowledge Cutoff
-
API Pricing (per 1M)
Input: $0.60 · Output: $2.50
Precision
Context
Workload Memory: 1938.09 GB
FP16 · 8K context
Base Weights: 1922.25 GB
KV Cache: 13.34 GB
Runtime & Graphs: 2.50 GB
16x NVIDIA H200 141GB
1938.1 / 2030.4 GB usable · TP 8 · PP 2
Determine requirements for your specific workload:
Upgrade to Pro
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.48.3
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashMLA |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Additional required packages
Install alongside standard runtime libraries:
pip install flash_attn
Rank
#98
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.811 | 27 |
Software Engineering | 0.53 | 31 |
Graduate-Level QA | 0.751 | 75 |
General Text | 1418 | 106 |
Web Development | auto 1322 | 114 |
Intelligence Index | auto 0.22 | 163 |
General Knowledge Reference | 0.895 | 🥉 3 |
Overall Rank
#98
Coding Rank
#110
Kimi K2-Instruct is an instruction-tuned 1T-parameter Mixture-of-Experts model from Moonshot AI engineered for agentic problem-solving and software development. It features advanced tool-calling and long-horizon reasoning across a 128K context window.
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
64
Key-Value Heads
64
Position Embedding
ROPE
RoPE Theta
50,000
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
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
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.