Atlas / Models / Moonshotai / MoonshotAI: Kimi K2 Thinking

MoonshotAI: Kimi K2 Thinking✓ Catalog verified

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 Kimi K2, it activates 32 billion parameters per forward pass and supports 256 k-token context windows. The model is optimized for persistent step-by-step thought, dynamic tool invocation, and complex reasoning workflows that span hundreds of turns. It interleaves step-by-step reasoning with tool use, enabling autonomous research, coding, and writing that can persist for hundreds of sequential actions without drift. It sets new open-source benchmarks on HLE, BrowseComp, SWE-Multilingual, and LiveCodeBench, while maintaining stable multi-agent behavior through 200–300 tool calls. Built on a large-scale MoE architecture with MuonClip optimization, it combines strong reasoning depth with high inference efficiency for demanding agentic and analytical tasks.

Input price
$0.47 /1M
Output price
$2.00 /1M
Context
131K
Modalities
text
Released
Nov 6, 2025
Tool calling
✓ Yes
Atlas signal
91/100
01

Overview

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 Kimi K2, it activates 32 billion parameters per forward pass and supports 256 k-token context windows. The model is optimized for persistent step-by-step thought, dynamic tool invocation, and complex reasoning workflows that span hundreds of turns. It interleaves step-by-step reasoning with tool use, enabling autonomous research, coding, and writing that can persist for hundreds of sequential actions without drift. It sets new open-source benchmarks on HLE, BrowseComp, SWE-Multilingual, and LiveCodeBench, while maintaining stable multi-agent behavior through 200–300 tool calls. Built on a large-scale MoE architecture with MuonClip optimization, it combines strong reasoning depth with high inference efficiency for demanding agentic and analytical tasks.

Access: available through the official Moonshotai API. Context window: 131K.

Pricing and metadata from the ModelsAtlas catalog.Last refreshed Aug 5, 2026
02

Specifications

API identifiermoonshotai/kimi-k2-thinking
ProviderMoonshotai
Model typeText LLM
Context window131K catalog
Input modalitiestext
Output modalitiestext
ReleasedNov 6, 2025
ModeratedNo
Architecture modalitytext->text
Supported parameters
frequency_penaltyinclude_reasoninglogit_biasmax_tokensmin_ppresence_penaltyreasoningrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_p
Source: provider documentation + catalog feedMethodology →

API defaults

Default parameters
{
  "top_p": null,
  "temperature": null,
  "frequency_penalty": null
}
03

Pricing

Live pricing components for moonshotai/kimi-k2-thinking as published in the catalog.

$0.47
Input /1M
$2.00
Output /1M
Per request
Per image
Pricing componentRaw unit priceNormalizedUnit
Prompt tokens$4.7e-7$0.47 / 1MPer input token
Completion tokens$0.000002$2.00 / 1MPer output token
Request feeN/AN/APer request
Image feeN/AN/APer image unit
Web search feeN/AN/APer search request
Source: catalog pricing feedCompare all pricing →Cheapest models →
04

Cost calculator

Estimate monthly spend from your own token volumes.

Scale / Volume

Estimated Total Cost

$0.64Calculated from current list pricing in the ModelsAtlas catalog.
05

Capabilities

CapabilityStatusWhat it means
Visual UnderstandingNot advertisedImage and document analysis support
Audio ProcessingNot advertisedSpeech and voice aligned flows
Tool Calling✓ SupportedSupports tools / function calling
Self-HostingNot advertisedDeploy outside managed APIs
Derived from catalog capability tags and supported parameters
06

Capability signals

Directional signals derived from model metadata and capability tags — not official benchmark submissions.

MoonshotAI: Kimi K2 Thinking
MMLU Signal (Reasoning)
97signal
Coding Signal (HumanEval proxy)
97signal
Math Signal (GSM8K proxy)
83signal
Science Signal (GPQA proxy)
87signal
Estimated from metadata — not an official benchmark runAll benchmarks →Methodology →
07

Quick start

Call MoonshotAI: Kimi K2 Thinking through an OpenAI-compatible client.

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_API_KEY"
)

response = client.chat.completions.create(
    model="moonshotai/kimi-k2-thinking",
    messages=[{"role": "user", "content": "Explain quantum physics."}]
)

print(response.choices[0].message.content)
08

Alternatives

Closest models by context window from a different provider.

09

Sources & attribution

Pricing and metadata are maintained in the ModelsAtlas catalog.

Capability bars use metadata tags — directional estimates onlyHow we source and verify data →
10

Frequently asked questions

How much does MoonshotAI: Kimi K2 Thinking cost?

$0.47 per 1M input tokens and $2.00 per 1M output tokens on the official Moonshotai API.

What is the context window of MoonshotAI: Kimi K2 Thinking?

MoonshotAI: Kimi K2 Thinking supports up to 131K tokens of context.

Does MoonshotAI: Kimi K2 Thinking support tool / function calling?

Yes, MoonshotAI: Kimi K2 Thinking supports tool / function calling — you can register tools and the model will emit structured tool calls.

How do I access MoonshotAI: Kimi K2 Thinking?

Use the official Moonshotai API with the model id `moonshotai/kimi-k2-thinking`. See the Quick start section above for code examples.