Skip to main content

Kimi K2

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 / output price
$0.57 / $2.30 per 1M
Context
131K tokens
Released
Jul 11, 2025
Agent setup

Paste into Claude Code, Cursor, or Codex

Help me set up the Respan gateway: fetch and follow https://www.respan.ai/docs/agent-setup.md. Use model "kimi-k2".

Read docs

Providers

3 providers, cheapest first, OpenRouter last, priced per 1M tokens after gateway discounts.

ProviderInputOutput
Novitanovita/kimi-k2
$0.57/M$2.30/M
Featherlessfeatherless/kimi-k2
$0.60/M$2.50/M
OpenRouteropenrouter/kimi-k2
$0.57/M$2.30/M

Is a provider down? OpenRouter status

Pricing

Averaged across 3 providers, in USD per 1M tokens after gateway discounts.

Average input
$0.58 per 1M
Average output
$2.37 per 1M
Average cache read
$1.00 per 1M

Cost by volume

Total cost from 0 to 1T tokens on the cheapest and most expensive provider.

Highest$0.60/MFeatherlessLowest$0.57/MNovita and OpenRouter

Throughput and latency

30-day averages from requests through the Respan gateway. Under each figure, every model on the gateway by its average, this one in blue.

Output speed

Higher is better

4.8tokens/sacross providers

Faster than 17% of the other 596 models on the gateway

Response latency

Lower is better

1.7sacross providers

Faster than 49% of the other 768 models on the gateway

Benchmarks

The Respan Index plus 19 published scores from 5 sources, each linked to where it was published.

17.6/ 100

Respan Index#146 of 176

176 ranked models. The shaded band is this model's 95% interval, 13.1 to 27.2. o3-mini and Qwen3.5 4B fall inside it, so the gap is within the margin of error.
Kimi K2 by category: Coding 5.0, Agents 12.5 (estimated), Reasoning 19.0, Knowledge & language 13.2, Human preference 47.7
  1. 146Kimi K217.6
  2. 145o3-mini17.7
  3. 147Qwen3.5 4B17.5

Published scores

BenchmarkScore
Aider-Polyglothuggingface.co60%
LiveCodeBench v6 (Aug 24 - May 25)huggingface.co53.7%
WeirdMLHåvard Tveit Ihle39.4%
TerminalBench (Inhouse Framework)huggingface.co30%
GSO Opt@1GSO2%

Compare Kimi K2: Kimi K2 vs Kimi K2.5, Kimi K2 vs Kimi K3.1 Pro, Kimi K2 vs Kimi K3, Kimi K2 vs Kimi K2.6, Kimi K2 vs Kimi K2.7 Code

Apps

Use Kimi K2 in the tools you already run. Each guide points the app at the Respan gateway.

Setup guides for more apps are in the integration docs.

Kimi K2 FAQ

Short answers, built from the data on this page.

Kimi K2 is a model from Moonshot, released on Jul 11, 2025. 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.

On the Respan gateway, Kimi K2 costs from $0.57 per 1M input tokens and $2.30 per 1M output tokens on Novita and OpenRouter, after gateway discounts. Averaged across its 3 providers, it costs $0.58 per 1M input tokens and $2.37 per 1M output tokens.

Kimi K2 is available through 3 providers: Novita, Featherless, and OpenRouter.

Kimi K2 accepts up to 131K tokens of context on its largest provider route.

Point any OpenAI-compatible SDK at https://gateway.respan.ai/v1 with your Respan API key and set the model to "kimi-k2". To set it up with a coding agent, copy the agent setup prompt at the top of this page.

Kimi K2 scores 17.6 out of 100 on the Respan Index v1.1, ranking #146 of 176 models. By category: Coding 5.0, Agents 12.5 (estimated), Reasoning 19.0, Knowledge & language 13.2, and Human preference 47.7. The Respan Index combines independent and lab-reported benchmark results into one score, using benchmark data collected by Respan.

Kimi K2 has 19 published benchmark scores, including Aider-Polyglot 60%, Tau2 retail 70.6%, and GPQA-Diamond 75.1%. Each score links to where it was published; the data is collected by Respan.

Certified where it counts.

Built to meet the security and privacy standards that enterprise and healthcare teams require.

ISO 27001

The internationally recognized standard for information security management.

SOC 2

Secure, compliant management of your data across all of our systems.

GDPR

Operated under GDPR, the world's strictest standard for data privacy.

HIPAA

HIPAA compliant, with a BAA available for healthcare teams.

Works with your entire stack

Use Respan with your favorite frameworks and tools.

AnthropicAI21 LabsxAIVercel AI SDK
LangChainOpenRouterAssemblyAICursor
LiteLLMGroqBasetenMem0
CohereCogneeLlamaIndexFireworks
DeepSeekLinkupHaystackTogether AI
PerplexityInference.netPostHogRubyLLM
SlackMastraAzure OpenAIMistral
AWS BedrockMoonshotHyperspellBAML
ModaInstructorGoogle GeminiParasail
ReductoZapierAgnoNebius
OpenAINovita AIReplicateSuperagent

Built for AI agents.
Break less.
Ship more.