Moonshot AI Releases Open-Source 1 Trillion Parameter 'Kimi K2'

Chinese AI startup Moonshot AI released its latest large language model “Kimi K2” as open-source on July 11, 2025. This Mixture-of-Experts (MoE) model boasting 1 trillion parameters is designed specifically for agent AI and code generation, garnering attention as part of Chinese AI companies’ strategy to expand their influence in the global market¹.

Kimi K2 features a massive scale with 1 trillion total parameters and 32 billion active parameters per token, achieving training on 15.5 trillion tokens using the proprietary “MuonClip” optimizer. The company reports achieving stable learning throughout the entire training process without any loss spikes².

Technical Specifications and Innovative Features

The technical characteristic of Kimi K2 lies in its “agent-oriented” architecture that stands apart from traditional chatbot-type AI. With native support for Model Context Protocol (MCP), it enables autonomous decomposition of complex tasks, execution of tool sequences, code writing and debugging, data analysis, and workflow orchestration³.

Benchmark Performance

The model demonstrates excellent performance across multiple benchmarks²:

SWE Bench Verified (software engineering tasks)
Tau2 (technical reasoning)
AceBench (agent capability evaluation)

In these benchmarks, Kimi K2 achieves state-of-the-art performance in the “non-thinking model” category, earning particularly high marks in frontier knowledge, mathematics, and coding⁴.

Implementation and Deployment Options

Moonshot AI provides two model versions catering to different use cases:

Kimi-K2-Base: Base pre-trained model for research and custom scenarios
Kimi-K2-Instruct: General instruction fine-tuned version for Q&A and agent tasks

In terms of technical specifications, it supports context lengths up to 128K tokens and is compatible with major inference engines including vLLM, SGLang, and ktransformers. API compatibility supports both OpenAI and Anthropic formats, facilitating easy integration into existing infrastructure².

Market Strategy and Open-Source Intent

This move by Moonshot AI can be understood as a response to market challenges the company is facing. The Kimi application, which ranked third in monthly active users as of August 2024, had fallen to seventh place by June 2025, impacted by competitor DeepSeek’s release of low-cost models¹.

Founded by Yang Zhilin in 2023, Moonshot AI is one of China’s prominent AI startups backed by internet giants including Alibaba. The company has adopted a strategy of releasing the model under a Modified MIT License to encourage widespread adoption and experimentation¹.

Pricing and Availability

The pricing for commercial use is announced at 4 yuan per million input tokens and 16 yuan per million output tokens. This represents competitive pricing in the Chinese market and is seen as strategic pricing aimed at expanding the developer community².

The trend of Chinese technology companies adopting open-source strategies is becoming established as a means to build global developer communities and expand international influence. The release of Kimi K2 will serve as an important example symbolizing this trend.

Detailed information about the model, documentation, and implementation examples can be found at Moonshot AI’s GitHub repository.