Moonshot

kimi-k2.5

LLM

Official description

Kimi K2.5 is the flagship open-source multi-modal language model product launched by Moonshot AI. It is the most intelligent model of Kimi to date, achieving the open-source state-of-the-art performance in Agent, code, visual understanding, and a series of general intelligent tasks. At the same time, Kimi K2.5 is also the most all-round model of Kimi to date. With its native multi-modal architecture design, it supports visual and text input, thinking and non-thinking modes, dialogue and Agent tasks. Its core positioning is "native visual intelligent agent engine", specifically designed for coding, office work, and complex task autonomous collaboration.
Performance breakthrough: Exhibits outstanding performance in the HLE, BrowseComp, and SWE-Verified three major intelligent agent benchmarks, with costs significantly lower than those of competitors.
Architectural innovation: The world's first open-source model that supports self-organizing "agent swarm" (Agent Swarm), capable of dynamically scheduling up to 100 sub-intelligent agents and executing up to 1,500 tool calls in parallel.
Coding revolution: The strongest open-source multimodal encoding model, capable of directly generating front-end interfaces with interactive animations from conversations or videos, achieving "what you see is what you get" visual programming.
Office efficiency improvement: End-to-end processing of 10,000-word documents, hundreds of pages of reports, financial models, and LaTeX formulas. Complex office tasks are reduced from hourly to minute-level processing time.
Efficiency leap: Compared to a single intelligent agent, the swarm mode reduces execution time by up to 80%, compresses key path steps by 3-4.5 times, and truly achieves "instant response".

View vendor docs ↗

Pricing

Input	$0.571 / 1M tokens
Output	$3 / 1M tokens

Specs

Context window: 256K tokens
Billing: Token-based
Available groups: default · svip · vip

Supported endpoints

openai/v1/chat/completions

36 production models, one endpoint.

36 production models,
one endpoint.