JAKARTA - Chinese artificial intelligence startup Moonshot launched Kimi K3, a model with 2.8 trillion parameters. Moonshot claims it is the world's largest open-weight AI model, with performance that is starting to approach the leading US model.
Kyodo News, quoted on Friday, July 17, said that the Kimi K3 was launched a month after the US government withdrew Anthropic's Fable and Mythos models from use due to security concerns.
The launch shows that the competition between Chinese and US AI is getting tighter. Moonshot, Z.ai, and MiniMax continue to release more powerful models at lower costs.
Kimi K3 is designed for advanced reasoning, long-term programming, and knowledge-based work. Moonshot calls it the first open-weight model that approaches three trillion parameters.
Parameters are elements in an AI model that help the system recognize patterns and generate answers.
Kimi K3 also has a one million token context window. Tokens are pieces of text that are processed by AI, such as words, word parts, or punctuation marks.
The feature allows the Kimi K3 to process and retain much more information in one command compared to previous generation models.
Moonshot claims that Kimi K3 is able to compete with Fable 5. The model is also said to outperform Opus 4.8, GPT-5.6 Sol, and GPT-5.5 in GPU kernel optimization.
The optimization is used to maximize the work of the AI hardware while reducing the waiting time when the model is running a task.
Third-party testing results also put Kimi K3 in the top ranks. Arena.ai put it in first place in the web interface creation benchmark.
Vals AI places Kimi K3 in second place overall, behind Fable 5 and above GPT-5.6 Sol.
Artificial Analysis judges its performance to be comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8, especially for complex tasks that require many stages.
Kyodo News also noted that Chinese AI companies are accelerating the launch of models as global competition intensifies.
Previously, Z.ai released GLM-5.2 which recorded scores close to the best closed model from the US in a number of tests.
The results have shaken the view of some Western analysts that China's AI model is still at least six months behind its American competitors.
MiniMax is also developing a model with 2.7 trillion parameters scheduled to launch as early as the third quarter of 2026.
The company is also preparing H3, a multimodal model that can process more than one type of data, such as text and images, as previously reported by Reuters.
The race towards models with trillions of parameters reflects the increasing need for autonomous systems capable of handling complex reasoning tasks.
A number of AI laboratories are also developing systems that are able to improve their own capabilities independently. This process is known as recursive self-improvement.
Before Kimi K3 came, Meituan's LongCat-2.0 and DeepSeek's V4-Pro led the Chinese AI industry with a model measuring 1.6 trillion parameters.
A number of other domestic competitors have also crossed the one trillion parameter threshold.
The open-weight model allows users to download, run, and customize the underlying system. This kind of model is different from the closed system controlled by the owner.
Moonshot said Kimi K3 brings two major changes to its architecture. The changes are claimed to improve computational efficiency and help the model complete long programming tasks with little human supervision.
The company, which is backed by Alibaba and Tencent, is also raising capital.
Bloomberg previously reported that Moonshot was seeking new funding of 2 billion US dollars with a valuation of around 30 billion US dollars ahead of a possible listing on the Hong Kong Stock Exchange.
The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)
