China Joins the AI Summit: Moonshot Unveils the 2.8 Trillion Parameter Kimi K3
Originally published in Turkish. Read the original →
China's Moonshot AI has unveiled Kimi K3, a new 2.8 trillion parameter model set to reshuffle the cards in the AI world. Standing up to the most advanced closed-source Western rivals such as Fable 5 and GPT-5.6 Sol, Kimi K3 delivers high performance at a much lower cost and will soon become fully open source.
Recently, the leadership seat in the AI race has been held by Western giants. OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 models were practically ruling the top alone with their superior performance. The enormous capability of these models had even led the US administration to impose restrictions on security grounds. But this week, Chinese AI startup Moonshot made a move that shook the belief that the West has a monopoly on the race to its core.
Moonshot AI announced Kimi K3, a new AI model that stands up to the most advanced closed-source models. According to the announcement on the company's official X account, this giant model has been released as open-weight, and the company said it will become fully open source as of July 27, 2026.
A New 2.8 Trillion Parameter Open-Source Giant
Kimi K3 has a massive architecture with exactly 2.8 trillion parameters. Although parameter count alone does not determine a model's quality, it is one of the most important indicators of how much engineering and data depth went into training it.
Described by Moonshot as "the largest open-weight AI model released to date", Kimi K3 was designed to perform especially well in the following areas:
* Advanced Reasoning: Analyzing complex problems and making logical inferences. * Long-Running Coding Tasks: Debugging and writing code end to end in large software projects. * Knowledge-Intensive Professional Work: Comprehensive data processing in fields such as academic research, financial analysis and law.
The model also has a large 1 million token context window. This lets it work on huge document sets, books or code blocks thousands of lines long while keeping them in memory within a single prompt.
Affordable Cost and the Flexibility of Self-Hosting
The biggest advantage of open-weight models like Kimi K3 is that, unlike the closed models of OpenAI or Anthropic, they give developers the freedom to download the model to their own infrastructure and customize it. This means a very significant long-term financial advantage, especially for large enterprises. Instead of paying external API usage fees, companies can run the model on their own servers, maximizing data privacy and lowering their costs.
This flexibility comes with a significant cost, though. Running a giant 2.8 trillion parameter model on local (on-premise) systems requires millions of dollars of investment in GPU hardware. For this reason, accessing it through cloud-based API services still stands out as the most sensible path for individual developers and mid-sized companies in the first phase.
Hardware Costs Are Falling with the Rubin Architecture
Although the cost of running models on local servers is high, new developments on the hardware side are changing this quickly. NVIDIA's new-generation Rubin architecture has started to shift the balance in AI servers.
Compared to previous-generation Hopper-based systems, the Rubin architecture:
1. Offers much higher memory capacity. 2. Multiplies memory bandwidth, increasing data processing speed. 3. Makes it possible to run models of similar size with far fewer GPU servers.
This hardware breakthrough significantly reduces operational expenses (OpEx) such as data center energy consumption, cooling costs and network infrastructure costs. Experts believe that thanks to new-generation hardware, hosting trillion-parameter models locally will become accessible to many more companies in the near future. That will keep increasing the appeal of open-source giants like Kimi K3 day by day.
