Kimi K2.5
Kimi K2.5 is a multimodal foundation model developed by Moonshot AI. It has a large-scale architecture with approximately 10 trillion parameters and a 256K context window.
Parameters
10000.0B
Context Window
256K
License
https://github.com/openai/gpt-2/blob/master/LICENSE
Release Date
2026-01-27
API Pricing
API pricing for this model is not yet available
Strengths
- ・Massive 10-trillion parameter scale
- ・Wide 256K context window
- ・Advanced multimodal support
Weaknesses
- ・Very large 595GB model size
- ・Closed licensing model
- ・Requires extremely high compute resources
Use Cases
- ・Ultra-large-scale data analysis and processing
- ・Understanding and summarizing long documents
- ・Integrating multimodal information
Deep Analysis
Architecture
MoE (~1T total, 32B active)
Open-source multimodal agentic model
Context Window
256K tokens
Training Data
~15T tokens
Mixed visual and text data
Release Date
January 27, 2026
License
Open-source (custom)
Modalities
Text + Vision
Native multimodal
Strengths
- ・Native multimodal agentic model with vision
- ・Beats Claude Opus 4.5 on several benchmarks
- ・Open-source with 15T token training
- ・Strong coding agent capabilities
- ・Visual agentic intelligence for complex tasks
Weaknesses
- ・Large model size requires significant compute
- ・Custom license may limit some commercial uses
- ・Relatively new with limited community tooling
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| Claude Opus 4.5 | - | - | - | Higher (closed) |
| GPT-5 | - | - | - | Higher (closed) |
| DeepSeek V3 | - | - | - | Comparable |
Kimi K2.5 is Moonshot AI's most powerful open-source model, combining native multimodal understanding with advanced agentic capabilities. Released January 2026, it outperforms Claude Opus 4.5 on several benchmarks and introduces visual agentic intelligence.
Sources
Analysis generated: 2026-05-24