The open model that was “checkmating” the giants: a year on, Kimi is almost three times bigger and still behind
In November 2025, people were sharing that a Chinese open-weights model, Kimi K2 Thinking, beat GPT-5 on some tests. Today Kimi K3, from the same family, is almost three times bigger, but its maker admits it does not match the best closed models.
On 10 November 2025, in ChatGPT-IA-edu, an article from WWWhatsnew was doing the rounds with a bold headline: Kimi K2 Thinking, from China’s Moonshot AI, “checkmates the giants of AI”. Newly launched, it had a trillion parameters, was released under a modified MIT licence and, according to the company’s own tests, beat GPT-5 and Claude Sonnet 4.5 at searching with tools. That an open model could measure itself against those from OpenAI, Google or Anthropic seemed a surprise.
In July 2026 Moonshot unveiled Kimi K3, with 2.8 trillion parameters and a context window of a million tokens; the weights reached Hugging Face on 27 July. But the announcement itself acknowledges that its overall performance still trails that of Claude Fable 5 and GPT 5.6 Sol.
What counts as open has also changed: K3 has its own licence, and anyone offering it as a service with more than 20 million dollars in annual revenue needs an agreement with Moonshot, according to VentureBeat. A year on, an open model measuring itself against closed ones is no longer news; catching them is not yet a reality.
The open model that was “checkmating” the giants: a year on, Kimi is almost three times bigger and still behind
In November 2025, people were sharing that a Chinese open-weights model, Kimi K2 Thinking, beat GPT-5 on some tests. Today Kimi K3, from the same family, is almost three times bigger, but its maker admits it does not match the best closed models.
On 10 November 2025, in ChatGPT-IA-edu, an article from WWWhatsnew was doing the rounds with a bold headline: Kimi K2 Thinking, from China’s Moonshot AI, “checkmates the giants of AI”. Newly launched, it had a trillion parameters, was released under a modified MIT licence and, according to the company’s own tests, beat GPT-5 and Claude Sonnet 4.5 at searching with tools. That an open model could measure itself against those from OpenAI, Google or Anthropic seemed a surprise.
In July 2026 Moonshot unveiled Kimi K3, with 2.8 trillion parameters and a context window of a million tokens; the weights reached Hugging Face on 27 July. But the announcement itself acknowledges that its overall performance still trails that of Claude Fable 5 and GPT 5.6 Sol.
What counts as open has also changed: K3 has its own licence, and anyone offering it as a service with more than 20 million dollars in annual revenue needs an agreement with Moonshot, according to VentureBeat. A year on, an open model measuring itself against closed ones is no longer news; catching them is not yet a reality.