Kimi K2.5
Moonshot AI's flagship agentic model via the Vercel AI Gateway. Unifies vision and text, thinking and non-thinking. 256k context. Supports image input. Served only by zero-data-retention providers. Does not support web search.
You always get the exact model you pick — we never silently route you to another.
About Kimi K2.5
Kimi K2.5 from Moonshot AI earns its reputation as the strongest open-source coding model by being built around vision from the ground up — its native MoonViT-3D encoder lets it convert mockups and screenshots directly into working frontend code, a workflow that text-only models simply cannot match. Developers report cutting three-month projects down to three weeks, and its 76.8% score on SWE-Bench Verified puts it ahead of most proprietary alternatives. Underneath sits a trillion-parameter mixture-of-experts architecture that keeps only 32 billion parameters active at inference, making the capability density unusually high relative to compute cost. An open-source release means the weights are auditable and fine-tunable. That said, real users consistently flag one genuine frustration: the model is verbose to a fault — ask a direct question and expect three paragraphs back. Independent analysis also puts its hallucination rate at 64%, so outputs need human review on factual tasks. At roughly 75–90% below OpenAI API pricing, it is a pragmatic trade-off for teams that need multimodal coding and document intelligence at scale and can build verification into their workflow.
Best for
- Design-to-code conversion: turning screenshots, mockups, or design files into functional, interactive frontend code
- Full-stack software development with strong multi-language support, especially complex UI and animation work
- Long-form document processing and research synthesis across very large inputs (up to ~150K tokens safely)
- Agentic and multi-agent task orchestration, including parallel research pipelines via Agent Swarm mode
- Document and image OCR extraction, competing with closed-source models on structured data tasks
Specifications
| Provider | Moonshot AI |
|---|---|
| Released | 2026-01 |
| Intelligence | High |
| Speed | Medium |
| Context window | 256,000 tokens |
| Input price | $0.60 / 1M tokens |
| Output price | $3.00 / 1M tokens |
| Request cost | 1 premium request |
| Plan tier | Premium |
| Input | Text, Image |
| Output | Text |
| Features | Function calling supported, Served through the Vercel AI Gateway (moonshotai/kimi-k2.5), Zero data retention and no prompt training |
| Model ID | kimi-k2p5 |
Frequently asked questions
Via OpenRouter it is $0.375 per million input tokens and $2.025 per million output tokens, with a 66% discount on cache hits. Alternative providers list $0.60 input / $3.00 output per million tokens.
The model supports up to 256K tokens, but accuracy degrades noticeably between 150K and 200K. Treat 150K as the practical working limit for tasks that require high reliability.
Frontend development, design-to-code workflows, and agentic automation — particularly tasks where visual inputs (images, screenshots) feed directly into code or structured output generation.
A high hallucination rate (flagged at 64% in independent testing), a tendency to over-explain simple answers, and reasoning instability when thinking mode and tool-calling are combined. Advanced mathematics is also a weak spot compared to GPT-4 variants.
Moonshot AI has since released K2.6 and K2.7-Code. K2.7-Code specifically targets coding workflows with reduced thinking-token overhead, so it may be preferable for pure coding tasks if latency matters.
Yes. Model weights and code are publicly available on Hugging Face and GitHub, enabling self-hosting, community audits, and fine-tuning — though self-hosting at this scale requires significant hardware.