A Chinese startup called Moonshot just launched Kimi K2 Thinkinga gigantic open model with a trillion parameters that has done something that seemed almost impossible: surpass the best proprietary models from companies like OpenAI, Google or Anthropic. If we thought that “Open Source” models could never compete with GPT-5, Gemini 2.5 Pro or Claude, we were wrong.
what has happened. This “AI laboratory” had already announced Kimi K2 in July with that gigantic size of one trillion parameters, but now they have released the “Thinking” version with that same size (32 billion active parameters, Mixture of Experts architecture). According to those responsible, the model is capable of maintaining stable use of agentic tools over between 200 and 300 sequential calls. Or what is the same: it can chain long sequences of actions autonomously and apparently without error. The best of all is not that: it is that it surpasses GPT-5 or Claude Sonnet 4.5 in various tests and costs much less than those models.
The benchmarks. Those responsible for Moonshot explained how Kimi K2 Thinking achieves the highest scores in Humanity’s Last Exam (general knowledge, 44.9%) and BrowserComp (agent browsers, 60.2%). He is almost at Claude’s level in the SWE software development test, and is also almost the best in another of those benchmarks, LiveCodeBench v6. It is true that in some tests still slightly behind of its “western” rivals, but the achievement is spectacular.
More benchmarks. Those responsible for Artificial Analysis have shown their first conclusions after evaluating it with various tests. Thus, they highlight its behavior in agentic tasks that simulate that the model is acting as a customer service agent. In this test it obtained 93% of the maximum, surpassing all its competitors by far (GPT-5 Codex High obtained 87%, for example). They will do more tests, but for now the prospects are fantastic.
And on top of that, cheap. On CNBC indicate that training the model cost $4.6 million, a ridiculous figure considering that training proprietary models like GPT-5 It cost about 500 million dollars according to estimates. Using the Kimi K2 Thinking API is also very affordable: $0.6 per million tokens in and $2.5 per million tokens out. GPT-5 Chat costs $1.25/10 respectively, while Claude Sonnet 4.5 costs $3/15 respectively.
The details. The model makes use of an INT4 quantization to improve its efficiency without compromising the precision and quality of its responses. Its context window—the “size” of the data we can enter when making prompts—is 256k, a relatively modest figure for large models but still notable. And as a good open model, we can download it to use locally… if we have a real monster at our disposal. The model weighs 594 GB, and for example joining two Mac Studio M3 Ultra It is possible to make it work locally relatively smoothly at about 15 t/s.
Alibaba is behindyes. Although the model is developed by an independent startup called Moonshot, this firm has been financially supported by Alibaba, which is becoming an absolute powerhouse in this field. Already not only conforms with developing its own models, which are outstanding (Qwen is the clear example), but is also financing the development of other models such as Kimi K2/Thinking.
China and its love for open AI models. During the last few months we have seen how China dominated in the field of open AI models —not “Open Source”—. The Asian giant has adopted an overwhelming philosophy with increasingly better models but which until now seemed to be several steps behind the large proprietary models of OpenAI, Anthropic or Google. This is no longer the case.
The race is lively. This achievement represents a new vote of confidence for the open models coming from Chinese companies. It is true that they are huge and that makes it very difficult to use them in practice by end users, but they present an interesting alternative for companies.
Image | idnaklss with Midjourney
In Xataka | There are many “internal” races within the greater AI race. And Alibaba is winning Open Source
GIPHY App Key not set. Please check settings