Hugging Face model card
Fields: tags, gated, license, summary, library name, pipeline tag
Kimi K2 Thinking
Kimi K2 Thinking is the latest, most capable version of open-source thinking model. Starting with Kimi K2, we built it as a thinking agent that reasons step-by-step while dynamically invoking tools. It sets a new state-of-the-art on Humanity's Last Exam (HLE), BrowseComp, and other benchmarks by dramatically scaling multi-step reasoning depth and maintaining stable tool-use across 200–300 sequential calls. At the same time, K2 Thinking is a native INT4 quantization model with 256k context window, achieving lossless reductions in inference latency and GPU memory usage.
- Context
- 262K
- Max output
- 100K
- Input
- $0.60
- Output
- $2.5
Recommended use cases
Strengths in this dataset
- 262,144-token context window
- text input
- 17 supported API parameters listed
Limits and caveats
- Provider behavior and pricing can change; verify the linked sources before production use.
Capability
Model-native facts
Model- Reasoning
- Supported
- Open weights
- Unknown
Provider endpoint facts
Provider endpoint- Tool calling
- Supported
- Structured output
- Supported
- Streaming
- Unknown
- Prompt cache
- Supported
- Batch
- Unknown
- Fine-tuning
- Unknown
Kimi K2 Thinking Provider pricing
Provider endpoint: moonshotai/kimi-k2-thinking
- Input
- $0.60 per 1M tokens
- Output
- $2.5 per 1M tokens
- Cached input
- $0.15 per 1M tokens
- Image output
- Unknown per 1M tokens
Sources and verification
Official model documentation
Fields: model identity
OpenRouter Models API
Fields: identity, description, modalities, context window, maximum output, pricing, supported parameters
Frequently asked questions
Answers are generated from the same sourced model and provider facts shown above.
Kimi K2 Thinking is the latest, most capable version of open-source thinking model. Starting with Kimi K2, we built it as a thinking agent that reasons step-by-step while dynamically invoking tools. It sets a new state-of-the-art on Humanity's Last Exam (HLE), BrowseComp, and other benchmarks by dramatically scaling multi-step reasoning depth and maintaining stable tool-use across 200–300 sequential calls. At the same time, K2 Thinking is a native INT4 quantization model with 256k context window, achieving lossless reductions in inference latency and GPU memory usage.
Model specifications and prices may vary by provider and change over time. AIToolly displays sources and verification dates so users can confirm critical details before production use.