Kimi K2.5 Model Benchmarks and Info
This article is a TL;DR of the Kimi K2.5 model benchmarks. For a more in-depth analysis of the model, check out Kimi's official blog post or the model on HuggingFace (respectively):
Model Architecture & Specs
Kimi K2.5 is a native multimodal agentic model built on a Mixture-of-Experts (MoE) architecture. It features a massive 1T parameter count with 32B active parameters per token, designed for high efficiency and reasoning depth.
| Feature | Specification |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Parameters | 1T Total / 32B Activated |
| Context Window | 256K Tokens |
| Vision Encoder | MoonViT (400M Params) |
| Layers | 61 (1 Dense) |
| Experts | 384 Total (8 selected per token) |
| Attention | MLA with 7168 Hidden Dimension |
Benchmark Highlights
Kimi K2.5 (Thinking Mode) was evaluated against top-tier models like GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro.
Reasoning & Knowledge
| Benchmark | Kimi K2.5 | GPT-5.2 (xhigh) | Claude 4.5 Opus | Gemini 3 Pro |
|---|---|---|---|---|
| AIME 2025 | 96.1 | 100 | 92.8 | 95.0 |
| HMMT 2025 (Feb) | 95.4 | 99.4 | 92.9 | 97.3 |
| GPQA-Diamond | 87.6 | 92.4 | 87.0 | 91.9 |
| HLE-Full (w/ tools) | 50.2 | 45.5 | 43.2 | 45.8 |
Coding & Agentic Search
| Benchmark | Kimi K2.5 | GPT-5.2 | Claude 4.5 Opus | Gemini 3 Pro |
|---|---|---|---|---|
| SWE-Bench Verified | 76.8 | 80.0 | 80.9 | 76.2 |
| LiveCodeBench (v6) | 85.0 | - | 82.2 | 87.4 |
| BrowseComp (Swarm) | 78.4 | - | - | - |
| DeepSearchQA | 77.1 | 71.3 | 76.1 | 63.2 |
Vision & Video
| Benchmark | Kimi K2.5 | GPT-5.2 | Claude 4.5 Opus | Gemini 3 Pro |
|---|---|---|---|---|
| MathVista (mini) | 90.1 | 82.8 | 80.2 | 89.8 |
| VideoMME | 87.4 | 86.0 | - | 88.4 |
| OCRBench | 92.3 | 80.7 | 86.5 | 90.3 |
Self-hosting
For self-hosting, Kimi K2.5 utilizes native INT4 quantization. It is compatible with the following inference engines (requires transformers >= 4.57.1):
- vLLM
- SGLang
- KTransformers
For full deployment guides and the "Kimi Vendor Verifier," refer to the official repository.