Kimi K2.5
Active Parameters
1T
Context Length
512K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
1708 KB/tok
License
Modified MIT License
Release Date
5 Feb 2026
Knowledge Cutoff
Oct 2025
API Pricing (per 1M)
Input: $0.50 · Output: $2.50
Hardware Sizing
Precision
Context
Workload Memory: 3681.53 GB
FP16 · 8K context
Base Weights: 3665.69 GB
KV Cache: 13.34 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
Cluster: 32x NVIDIA H200 141GB
3681.5 / 4060.8 GB usable · TP 8 · PP 4
Determine requirements for datacenter hardware:
Upgrade to Pro
Architecture Diagram
Evaluation Benchmarks
Rank
#66
| Benchmark | Score | Rank |
|---|---|---|
Professional Knowledge | 0.871 | 5 |
Software Engineering | 0.71 | 14 |
StackUnseen | 0.649 | 18 |
Graduate-Level QA | 0.876 | 33 |
General Text | auto 1450 Standard 1431 | 68 93 |
Web Development | auto 1436 Standard 1406 | 70 79 |
Coding Index | 0.47 | 93 |
Intelligence Index | auto 0.23 Standard 0.19 | 127 163 |
Rankings
Overall Rank
#66
Coding Rank
#79
About Kimi K2.5
Kimi K2.5 is a 1T-parameter native multimodal Mixture-of-Experts model from Moonshot AI engineered for complex reasoning and agent swarm orchestration. It supports up to 100 parallel sub-agents and visual coding workflows across a 256K context window.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
64
Key-Value Heads
64
Position Embedding
Absolute Position Embedding
RoPE Theta
50,000
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
968.0B
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.