Moonshot AI
Kimi-K3
About model
Moonshot AI's newest flagship MoE model, extending Kimi-2.7 with stronger reasoning and agentic tool-use performance across a 1M context window
It takes text as input and returns text. It works with a context window of 1M tokens and can return up to 1M tokens in a single response.
Calls go through the SCX gateway on an OpenAI compatible endpoint, so moving an existing integration across means changing the base URL and the model name, and nothing else.
1M
Tokens the model can attend to in a single request
79.2
Overall score, ranked 5 on the public leaderboard
1M
Tokens the model can return in one response
Model key capabilities
Reasoning
Works through a problem step by step before answering, which lifts accuracy on multi-step maths, logic and planning tasks.
Tool calling
Returns structured arguments for the tools you define, using the same schema as the OpenAI API, so existing agent frameworks work unchanged.
JSON mode
Constrains output to valid JSON, so responses can be parsed without a repair step.
Streaming
Streams tokens as they are generated, so an interface can show output before the response completes.
Benchmarks
Scores on LiveBench, a contamination-free benchmark refreshed every six months. Higher is better.
- Reasoning
- 90.7
- Coding
- 81.4
- Agentic Coding
- 62.2
- Mathematics
- 84.4
- Data Analysis
- 78.7
- Language
- 85.5
- Instruction Following
- 71.4
Source: LiveBench 2026-06-25 release.
Quick start
curl https://api.scx.ai/v1/chat/completions \
-H "Authorization: Bearer $SCX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Kimi-K3",
"messages": [{ "role": "user", "content": "Hello" }]
}'