Kimi K2.5

// Overview

Kimi K2.5 is Moonshot AI's open-weight flagship model, released on January 27, 2026 as the multimodal successor in the company's Kimi K2 line. It is a Mixture-of-Experts model with 1 trillion total parameters but only 32 billion active per token, which keeps inference cost low relative to its size. Unlike the original text-only Kimi K2, K2.5 is natively multimodal: it was built by continued pretraining over roughly 15 trillion mixed visual and text tokens and pairs the language backbone with a 400M-parameter MoonViT vision encoder.

K2.5 accepts text, images, and (experimentally) video, with a 256K-token context window. Moonshot positions it as a 'visual agentic' model aimed at real work, exposing four operating modes: Instant for quick answers, Thinking for step-by-step reasoning, Agent for research and content creation, and an Agent Swarm mode that can coordinate up to 100 sub-agents across many parallel steps for large-scale research, long-form writing, and batch tasks.

The weights are published on Hugging Face under a Modified MIT License that permits commercial use, so K2.5 can be self-hosted. It is also served through Moonshot's own Kimi API platform and through third parties including Amazon Bedrock and OpenRouter. Note that Moonshot has since shipped newer K2.6 and K2.7 models, so K2.5 is now an earlier release in the line rather than the current flagship.

// Benchmarks

1. SWE-bench Verified76.8%
2. SWE-bench Multilingual73%
3. SWE-bench Pro50.7%
4. Terminal-Bench 2.050.8%
5. LiveCodeBench v685%
6. AIME 202596.1%
7. GPQA-Diamond87.6%
8. MMLU-Pro87.1%
9. Humanity's Last Exam (full)30.1%
10. VideoMMMU86.6%
11. VideoMME87.4%
12. MMMU-Pro78.5%
13. OCRBench92.3%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Standard list price on Moonshot's Kimi API. Third-party providers vary; OpenRouter lists about $0.375 input / $2.025 output per 1M tokens. Self-hosting the open weights is also an option.

Pricing source ↗

// Strengths

• Strong agentic coding: 76.8% on SWE-bench Verified and 73.0% on SWE-bench Multilingual, competitive with leading proprietary models
• Open weights under a Modified MIT License that allows commercial use and self-hosting
• Cost-efficient MoE design that activates only 32B of 1T parameters per token
• Native multimodal vision (built on a MoonViT encoder) with strong document, OCR, and video-understanding scores
• Agent Swarm mode that parallelizes work across up to 100 coordinated sub-agents
• Large 256K-token context window for long documents and codebases

// Best for

• Agentic software engineering: multi-file bug fixes, refactors, and repo-level tasks
• Visual coding and turning screenshots or design images into working code
• Long-document analysis, OCR, and structured extraction from PDFs and scans
• Large-scale parallel research and long-form writing via Agent Swarm
• Self-hosted deployment for teams needing on-prem or private inference
• Multilingual coding and reasoning across many programming languages

// How to access

// Kimi K2 — every version

The full lineage of the Kimi K2 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.