AI MODELS
A Chinese open-source model just took the top spot
Chinese start-up Moonshot AI has launched a new reasoning version of its open-source model, Kimi K2 Thinking, which has scored higher than OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 on several major tests.
The model is now available on Kimi.com and through an API for developers.
A GitHub post from the research team says Kimi K2 Thinking reached new top results in reasoning, coding, and agent-style tasks, suggesting Chinese AI groups are closing the performance gap with leading US models.
Key results include:
44.9% on Humanity’s Last Exam, ahead of GPT-5 and Claude Sonnet 4.5
60.2% on BrowseComp, which measures web-based research
56.3% on Seal-0, focused on real-world research questions
Consultancy Artificial Analysis also ranked it first on Tau-2 Bench Telecom with 93% accuracy.
Who’s sweating?
The model uses a Mixture-of-Experts design with 1 trillion parameters, and its API access is estimated to be 6–10x cheaper than similar US systems.
Moonshot AI says performance improved thanks to a “model-as-an-agent” training method, helping it use different tools during complex reasoning.
Commentators say these results show open-source systems are getting closer than ever to high-end closed-source models.
In brief:
Kimi K2 Thinking topped well-known models in several benchmark tests
Pricing is reported to be significantly lower
Analysts say the performance gap between open-source and closed-source models is narrowing
Industry voices have called this a notable moment for open-source AI, pointing to stronger competition around price and capability.
1 trillion parameters is crazy… my brain crashes at 3 tabs bro. - MG