Moonshot releases Kimi K2 Thinking model with native INT4 quantization

Moonshot has released the Kimi K2 Thinking model card, detailing its capabilities as an open-source thinking agent that reasons step-by-step while dynamically invoking tools. The model sets a new state-of-the-art on benchmarks including Humanity's Last Exam (HLE) and BrowseComp by scaling multi-step reasoning depth.

• Kimi K2 Thinking is a native INT4 quantization model featuring a 256k context window, achieving lossless reductions in inference latency and GPU memory usage.
• It utilizes Quantization-Aware Training to achieve a 2x speed-up in low-latency mode while maintaining stable tool-use across 200–300 sequential calls.
• The model demonstrates coherent goal-directed behavior over long horizons, surpassing prior models that degrade after 30–50 steps.

The release provides instructions for using the model with libraries like Transformers and inference providers such as vLLM and SGLang.