Moonshot AI has released the open-weights Kimi K2 Thinking and Kimi K2 Thinking Turbo, trillion-parameter reasoning models designed to outperform proprietary LLMs in agentic tasks. These models execute hundreds of tool calls sequentially while interleaving reasoning steps, allowing them to adjust plans based on interim results.
- Architecture: Mixture-of-experts transformer with 1 trillion total parameters and 32 billion active per token.
- Performance: Achieved state-of-the-art results on the τ²-Bench Telecom agentic benchmark (93% accuracy) and Humanity's Last Exam with tools enabled (44.9%).
- Efficiency: Fine-tuned at INT4 precision, reducing file size to 594 GB and doubling output speed compared to full precision.
- Availability: Weights available under a modified MIT license for noncommercial and commercial uses up to 100 million monthly active users or $20 million monthly revenue.
Agentic applications benefit from the model's ability to reason across many tool calls without human intervention, while INT4 precision enables deployment on less expensive hardware.