Moonshot AI Unveils World's Largest Open-Source AI Model

Moonshot AI, a Beijing-based AI startup, released its large language model 'Kimi K3' in July 2025. The model features 2.8 trillion parameters—numerical values that serve as the knowledge and decision-making criteria acquired during AI training—and the company positions it as the world's largest open-source AI model. The complete model weights are scheduled for public release on July 27, and trial access is currently available by creating an account at kimi.com.

Moonshot AI is a company backed by Alibaba and has been leading AI development domestically in China. However, over the past 18 months, the company's market presence had declined due to DeepSeek's rapid rise. The release of Kimi K3 is seen as a major move to reverse that situation. The timing of this release coincides with the upcoming World Artificial Intelligence Conference (WAIC) to be held in Shanghai in 2026, and it is also noteworthy in terms of demonstrating the company's presence to the international community.

From a technical perspective, Kimi K3 reportedly has approximately 75% more parameters than DeepSeek's V4 Pro. The model features a context window of up to 1 million tokens (a unit of length allowing processing of large amounts of text or conversations at once), multimodal capabilities for directly understanding images, and a mode called 'thinking mode' that enables continuous reasoning. In terms of architecture, two technologies developed in-house by Moonshot AI—'Kimi Delta Attention' and 'Attention Residuals'—are employed, both of which have been previously published as open research on GitHub.

API (connection interface for linking with external services) pricing is set at 3 dollars per 1 million input tokens and 15 dollars per 1 million output tokens. Cached input tokens are reduced to 0.30 dollars per 1 million. Additionally, a campaign offering vouchers worth up to 30% is currently running through August 12 for API usage exceeding 1,000 dollars. For developers, compatibility with OpenAI's SDK is ensured, making it easy for engineers already developing with OpenAI or Anthropic tools to switch. According to Moonshot AI, the model is designed for straightforward migration.

In benchmarks—standard tests for comparing AI model performance—Moonshot AI claims that Kimi K3 can compete at a level comparable to currently top-tier commercial models such as Anthropic's Claude and OpenAI's GPT. The prospect of an open-source model directly competing with closed commercial systems has the potential to overturn conventional wisdom. According to reports by Xinhua News Agency, Moonshot AI executives explained the significance of the parameter count as follows: 'Similar to neural connections in the human brain, having close to 3 trillion parameters enables accumulation of more knowledge and patterns, deeper thinking, and more accurate answers.'

This move is positioned as accelerating a structural shift in AI model development competition between 'open' and 'closed' approaches. The public release of high-performance models in freely usable and modifiable form significantly lowers the barrier to access for research institutions and small-to-medium-sized developers. Conversely, as the gap in performance narrows, the competitive environment for Western companies providing commercial models will also shift. Going forward, external validation by researchers following the actual weights release on July 27 will be a key focus in determining how much the company's performance claims can be substantiated.