Chinese startup Beijing Moonshot AI Co. Ltd. Thursday released a new open-source artificial intelligence model, named Kimi 2 Thinking, that displays significantly upgraded tool use and agentic capabilities compared to current models.

The company said the new model was built as a “thinking agent,” capable of reasoning step-by-step while using software tools such as search, advanced calculations, data retrieval and third-party services. The model itself can reportedly execute 200 to 300 sequential tool calls without human intervention and reason across their use.

The model weighs in at 1 trillion parameters, a fairly large size for an open-source model. A source close to the project told CNBC it only cost $4.6 million to train. DeepSeek’s V3 model, also the product of a Chinese company and the foundation for the R1 reasoning model, reportedly cost $5.6 million.

When AI developers talk about parameter counts, they’re referring to the individual adjustable values within a model that determine how it processes and understands information. In essence, parameters are the neurons of a digital brain — the more of them a model has, the more complex patterns it can recognize and reason about. However, size isn’t everything; efficient training methods and specialized model architecture often matter more than parameter count.

Compared with the hundreds of millions or even billions of dollars that companies such as OpenAI Group PBC and Anthropic PBC spend to develop and train their frontier models, Moonshot’s figures highlight a growing arms race to discover cheaper, faster paths to advanced AI development.

It’s also noteworthy that DeepSeek’s R1 model has 673 million parameters, meaning that K2’s trillion-parameter scale at low training cost represents a continuing innovation in algorithmic and economic efficiency.

This release comes shortly after Nvidia Corp. Chief Executive Jensen Huang commented that China is primed to beat the United States in the AI race. He cited patchwork regulations between U.S. states, in contrast to the more unified regulatory approach by China and the country’s energy subsidies. Huang has been a strong critic of export controls on Nvidia’s AI chips to China, arguing that the country will eventually develop alternatives.

“Chinese AI researchers will use their own chips,” the Financial Times quoted the executive in May. “They will use the second-best. Local companies are very determined, and export controls gave them the spirit, and government support accelerated their development. Our competition is intense in China.”

Despite its low training cost, Moonshot claims that Kimi K2 Thinking outperforms many rival closed-source models on modern benchmarks. The model achieved a 43% score on Humanity’s Last Exam, a benchmark of 3,000 graduate-level reasoning questions designed to measure complex problem-solving beyond simple recall. The company claimed these results notably exceed the performance of OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5.

For software engineers, the new model delivered improvements in HTML, React and detailed front-end programming tasks. Users could use it to translate prompts into fully functional, responsive products using its agentic coding capabilities.

According to Moonshot, the model employs long-horizon planning combined with adaptive reasoning to break down tasks, use tools, and refine hypotheses before arriving at a final answer.

On the deployment side, the company said K2 natively supports INT4 inference, a compact quantized form of the model that reduces memory footprint and enables it to run on less powerful hardware. Low-bit quantization improves efficiency and latency for large-scale deployments; however, reasoning models such as K2 perform multiple cognitive passes, meaning high compression can degrade accuracy.

Moonshot said it overcame that with Quantization-Aware Training, or QAT, during the post-training phase, applying INT4 weights to the mixture-of-experts components. That allows K2 Thinking to support INT4 with around double the speed improvements while still achieving state-of-the-art accuracy and performance.

Image: SiliconANGLE/Bing Image Creator