Felo API PlatformFelo API Platform
Back to models
Moonshot AImessages

MoonshotAI: Kimi K2.5 API

Kimi K2.5 is Moonshot AI's open-weight native multimodal model with a 262K-token context, agent swarms and frontier-level coding, vision and agentic performance.

Kimi K2.5 is an open-weight native multimodal model from Moonshot AI that understands both text and images. Built on Kimi-K2-Base through continued pretraining on about 15 trillion mixed vision and text tokens, it uses a Mixture-of-Experts architecture with 1.04 trillion total parameters and 32B active parameters per inference, plus a 400 million parameter MoonViT vision encoder. It offers a thinking mode for step-by-step analysis and an instant mode for fast responses, and supports tool calling through an OpenAI/Anthropic-compatible API.

Its agentic capabilities are a defining feature: Agent Swarm decomposes complex tasks into parallel subtasks handled by dynamically instantiated domain-specific agents, coordinating up to 100 subagents and executing up to 1,500 tool calls. Coding with Vision turns visual specifications such as UI designs and video workflows into code, and the model autonomously coordinates tools for visual data processing. Released under a modified MIT license permitting commercial use, it runs on inference engines including vLLM, SGLang and KTransformers.

The model suits multi-agent task coordination at scale, complex data processing pipelines that scrape, clean, analyze and visualize data in parallel where dependencies allow, code generation from visual specifications, and agentic workflows executing 200 to 300 tool calls autonomously. Kimi Code CLI is recommended as the agent framework for optimal performance. Among open-weight models it took first place on 13 of 17 Vals AI benchmarks, though Claude Opus 4.5 remains the leader on SWE-Bench for maximum code generation accuracy.

Updated:

MODEL ID

kimi-k2.5
Context
262.144K
Max output
235.929K
Recommended API
messages
Input
text / image
Output
text

Highlights

  • Native multimodal text and image understanding
  • Agent Swarm with up to 100 subagents
  • Up to 1,500 tool calls
  • Thinking and instant modes
  • Open weights under modified MIT license
  • Runs on vLLM, SGLang and KTransformers

Providers

2 providers

Compare API prices across providers. Select a provider to view all charges and conditions.

Provider · Pricing details
Uptime
$0.49$2.5$0.2
$0.6$3Not published

What $10 buys

Input
20.4M tokens
Output
4M tokens

Estimated from current input and output prices, before discounts.

Benchmarks

  • Vals Index

    Ranked first among open-weight models

    #1
  • Vals Multimodal Index

    Ranked first among open-weight models

    #1
  • SWE-bench Verified

    Ranked first among open-weight models

    #1
  • Terminal-Bench 2.0

    Ranked first among open-weight models

    #1
  • Finance Agent

    Ranked third among all models

    #3
  • SAGE

    Ranked third among all models

    #3
  • HLE

    Achieved with tool use

    50.2%
  • CaseLaw

    Its lowest rank, within 10% of first place

    #16

Scores as published by the model provider.

Reasoning

Enabled by default
Yes
Always enabled
No

Quick Start

Drop-in code to call this model with Felo API Platform compatible APIs. Start in the Playground when you want to test prompts before writing code.

1

Get your API key

Create a key from your Felo API Platform account and set it as an environment variable.

export FELO_API_KEY="YOUR_API_KEY"
2

Make your first request

Use this model ID with the Felo API Platform endpoint. Optional request headers such as X-Request-Id can help troubleshooting.

kimi-k2.5

This model is reasoning-capable. If your selected protocol and model route expose reasoning controls, preserve the returned reasoning metadata when continuing a conversation.

const response = await fetch("https://openapi.felo.ai/api/v1/messages", {
  method: 'POST',
  headers: {
    Authorization: `Bearer ${process.env.FELO_API_KEY}`,
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
  "model": "kimi-k2.5",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Hello"
    }
  ]
}),
});
if (!response.ok) throw new Error(await response.text());
console.log(await response.json());

Using third-party SDKs

Use compatible SDKs by setting baseURL to https://openapi.felo.ai/api/v1. Use the Anthropic-compatible base URL for Claude-style clients. https://openapi.felo.ai/api.

3

Enable streaming

Add stream: true to receive responses as server-sent events.

curl -N https://openapi.felo.ai/api/v1/messages \
  -H "Authorization: Bearer $FELO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "kimi-k2.5",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Hello"
    }
  ],
  "stream": true
}'

Endpoint

Felo API Platform exposes compatible model-calling surfaces. Pick the surface your app or agent client already supports.

SurfaceMethodURL

Messages

Anthropic-compatible messages for Claude-style clients.

POST

https://openapi.felo.ai/api/v1/messages

Parameters

NameTypeDefaultDescription
modelstringkimi-k2.5Public model ID to call.
max_tokensinteger—Maximum generated tokens.
messagesarray—Conversation messages for Chat Completions.
seed———
stop———
min_p———
toolsarray—Tool calling definitions following compatible schema.
top_k———
top_p———
logprobs———
reasoning———
logit_bias———
temperaturenumber—Controls response variety when supported.
tool_choicestring or object—Controls whether a tool should be called.
top_logprobs———
response_formatobject—Request structured output when supported.
presence_penalty———
frequency_penalty———
include_reasoning———
repetition_penalty———
structured_outputs———

Frequently asked questions

What is MoonshotAI: Kimi K2.5?

Kimi K2.5 is an open-weight native multimodal model from Moonshot AI that understands both text and images. Built on Kimi-K2-Base through continued pretraining on about 15 trillion mixed vision and text tokens, it uses a Mixture-of-Experts architecture with 1.04 trillion total parameters and 32B active parameters per inference, plus a 400 million parameter MoonViT vision encoder. It offers a thinking mode for step-by-step analysis and an instant mode for fast responses, and supports tool calling through an OpenAI/Anthropic-compatible API. Its agentic capabilities are a defining feature: Agent Swarm decomposes complex tasks into parallel subtasks handled by dynamically instantiated domain-specific agents, coordinating up to 100 subagents and executing up to 1,500 tool calls. Coding with Vision turns visual specifications such as UI designs and video workflows into code, and the model autonomously coordinates tools for visual data processing. Released under a modified MIT license permitting commercial use, it runs on inference engines including vLLM, SGLang and KTransformers. The model suits multi-agent task coordination at scale, complex data processing pipelines that scrape, clean, analyze and visualize data in parallel where dependencies allow, code generation from visual specifications, and agentic workflows executing 200 to 300 tool calls autonomously. Kimi Code CLI is recommended as the agent framework for optimal performance. Among open-weight models it took first place on 13 of 17 Vals AI benchmarks, though Claude Opus 4.5 remains the leader on SWE-Bench for maximum code generation accuracy.

How is usage billed?

Your bill depends on the provider you use, its prices and your usage. Longer inputs or other conditions may have different prices.

What input modalities does Kimi K2.5 support?

It accepts text and image input and produces text output as a native multimodal model.

How long is its context window?

Kimi K2.5 supports a context of 262K tokens.

What is Agent Swarm?

Agent Swarm decomposes complex tasks into parallel subtasks run by dynamically instantiated domain-specific agents. It can coordinate up to 100 subagents and execute up to 1,500 tool calls.

What are the thinking and instant modes?

Thinking mode provides step-by-step analysis, while instant mode is tuned for fast responses. The recommended temperature is 1.0 in thinking mode and 0.6 in instant mode, with top_p 0.95.

How does it differ from Kimi K2 Thinking?

The previous Kimi K2 Thinking did not support multimodal input, whereas Kimi K2.5 is a native multimodal model that understands both text and images.

What license is it released under?

It is released under a modified MIT license that permits commercial use.

Related models