Call it déjà vu: A Chinese AI firm has set the industry abuzz by doing more with less.
Driving the news: New tests show that the new variant of Moonshot AI’s Kimi K2 model outperformed heavy hitters like OpenAI’s GPT-5 in several benchmark tests, setting records in areas like coding and agentic capabilities. It’s also been praised for its creative writing abilities.
- Like other open-source Chinese models, Kimi was made relatively cheaply. The CEO confirmed the model uses older and less powerful Nvidia chips.
How it works: Unlike many LLMs, Kimi relies heavily on interactions with external tools to answer queries. It was also designed to not use all of the variables it learned during training to solve a given problem, instead using only “relevant parts” of the model.
Why it matters: Kimi isn’t the best at everything, but it does excel at the run-of-the-mill chatbot tasks most people (and many businesses) are actually using. Moonshot’s ability to achieve this, through deliberate focus and open-source sharing, calls into question the viability of costly, closed-off training.
- "It’s clear that the surge of open models should make the closed labs sweat,” Nathan Lambert, a top AI researcher, wrote on his Interconnects Substack.
Big picture: Nvidia CEO Jensen Huang walked back recent comments about China winning the AI race, but he may have had a point. The emergence of Moonshot and other fledgling firms proves DeepSeek was no fluke.—QH