Moonshot AI Launches World’s Largest Open-Source AI Model to Challenge US Leaders
China’s Moonshot AI has released the world’s largest open-weight AI model, Kimi K3, with 2.8 trillion parameters, escalating the race with major US AI startups. Large language model Kimi K3 is built for advanced reasoning, coding and long-context processing.
Unveiled on July 17th, the Kimi K3 model has 2.8 trillion parameters and was developed by the Chinese AI startup Moonshot. According to the company, this is the biggest open-weight AI system in the world. Further, it has achieved results that are comparable to those of Anthropic's frontier Fable model, a US AI creative giant.
A month after the US government abruptly removed Anthropic's Fable and Mythos models on security concerns, China's open AI ecosystem is rapidly catching up to the most advanced US systems, as demonstrated by the debut. Moonshot, Z.ai, and MiniMax are among the companies launching models that are getting progressively more powerful at drastically reduced prices. This goes against the Western belief that Chinese programmers are months behind their American counterparts, which has been held for a long time.
Loaded Features of Kimi K3
Kimi K3 is the first open-weight model to approach the 3-trillion-parameter mark, according to Moonshot. Additionally, it is built for knowledge work, long-horizon coding, and complex thinking. The model's 1 million-token context window makes it capable of processing and retaining much more data in a single prompt compared to previous generations. According to the business, Kimi K3 had superior GPU kernel optimisation compared to (OpenAI's) Opus 4.8, GPT 5.6 Sol, and GPT 5.5, and it was competitive with Fable 5 (with fallback).
The term describes methods that reduce latency while maximising the use of AI hardware. In a test that evaluated web interface-building capabilities, Kimi K3 was ranked top by Arena.ai, second overall by Vals AI, and ahead of both Fable 5 and GPT-5.6 Sol by Vals AI. On tests measuring complicated, multi-step tasks, the model produced performance comparable to OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8, according to Artificial Analysis.
Chinese Firm Heading Ahead in AI Race
As the worldwide AI arms race heats up, Chinese AI companies are speeding up the cycles in which they produce new models. The change comes after Z.ai's GLM-5.2 made its debut, shocking industry watchers with benchmark test scores that were close to those of the best US closed-source models. This disproves the widely held belief among Western experts that China's AI models were six months behind schedule. Additionally, MiniMax, a Hong Kong-listed company, is working on its own 2.7-trillion parameter model, which might be unveiled as early as the third quarter of 2026.
In addition, their multimodal model H3, which is at the edge of innovation, is set to be released soon. The increasing demand for autonomous systems that can handle difficult reasoning tasks is reflected in the race for trillion-parameter systems. Some of the top AI research centres are also working on systems that can improve themselves automatically; this is known as recursive self-improvement. With 1.6 trillion total parameters, Meituan's LongCat-2.0 and DeepSeek's V4-Pro were China's leading AI models before Kimi K3 was released, while numerous other domestic competitors have also surpassed the trillion-parameter mark.
