MoonshotAI: Kimi K2 Thinking
Kimi K2 Thinking is Moonshot AI’s most advanced open reasoning model to date, extending the K2 series into agentic, long-horizon reasoning. Built on the trillion-parameter Mixture-of-Experts (MoE) architecture introduced in...
Intelligence overview
262.1K tok
Context window
98.3K tok
Max output
What MoonshotAI: Kimi K2 Thinking can do.
Reasoning
Extended reasoning
Works through problems step by step before answering, trading latency for accuracy on harder tasks.
Development
Tool calling
Calls external functions mid-response, then continues reasoning from the result.
Structured output
Returns responses constrained to a JSON schema, ready to parse without cleanup.
Streaming
Streams tokens as they're generated instead of waiting on the full response.
Performance
Design Arena — ranked #91 across 1 categories
Pay for what you use.
Every model in UltraGPT runs behind one flat subscription — this is the underlying per-token cost the catalog tracks.
$0.6 / 1M tok
$2.5 / 1M tok
Built for serious work.
Build
Generate and refactor code, with the ability to run and verify tool calls and return structured, parseable output.
Think
Works through multi-step problems with extended reasoning before it answers.
Research
Holds up to 262.1K tokens of source material in context without losing the thread.
Automate
Chains tool calls across multi-step workflows, reasoning between each one.
Ask MoonshotAI: Kimi K2 Thinking anything.
A preview of the real UltraGPT composer — the model you see here is the model you get.
Continues in the UltraGPT app — one login, every model.
How MoonshotAI: Kimi K2 Thinking stacks up.
The closest models in the catalog by provider, capability and score — real entries, pulled from the same directory.
Technical details.
Specification
- Context window
- 262.1K tokens
- Max output
- 98.3K tokens
- Input modalities
- Text
- Output modalities
- Text
- Release date
- Nov 6, 2025
- Streaming
- Supported
- Tool calling
- Supported
- JSON mode
- Supported
More from Moonshot AI, and everything else.
This model is one entry in the full UltraGPT catalog — browse the rest, compare pricing and context windows, or filter by capability.
