China’s 2.8T-Parameter Model Raises the Stakes

China’s AI sector has a new heavyweight. Beijing-based Moonshot AI launched Kimi K3 on July 17, 2026, describing the 2.8-trillion-parameter system as the world’s largest open-source AI model.

Kimi K3 targets long-horizon coding, reasoning, and knowledge work. Its release gives companies another potential alternative to proprietary systems from OpenAI and Anthropic, while adding momentum to China’s open-model ecosystem. The model’s scale, however, could make private deployment impractical for many businesses across Asia-Pacific.

Kimi K3 approaches the 3-trillion-parameter mark

Reuters reported that Kimi K3 is the first open-weight model to approach the 3-trillion-parameter threshold. Open-weight models allow organizations to download and customize the underlying system, while proprietary models remain controlled by their developers.

Moonshot said Kimi K3 includes a 1-million-token context window, which could allow it to process large codebases, documents, and extended workflows in a single request.

Recommended reading

OpenAI’s $70 ChatGPT Basketball Tests Brand Power

According to Reuters, Moonshot said Kimi K3 “performed competitively with Fable 5” and substantially outperformed Anthropic’s Opus 4.8, OpenAI’s GPT-5.6 Sol, and GPT-5.5 in GPU kernel optimization tests. The South China Morning Post noted that Moonshot also acknowledged Kimi K3 still trailed the strongest proprietary models in overall performance.

Independent evaluations provided additional support for some of Moonshot’s claims. Reuters said Arena.ai ranked Kimi K3 first in a test focused on building web interfaces, while Vals AI placed it second overall behind Anthropic’s Claude Fable 5.

China’s model race puts pressure on rivals

Kimi K3 arrived only weeks after competing releases from Chinese developers including Z.ai, DeepSeek, and MiniMax. Shares of Z.ai and MiniMax fell sharply in Hong Kong after Moonshot’s announcement, highlighting how quickly investor expectations are shifting around model performance.

Bank of America analysts said the launch showed that Chinese developers could continue improving despite restrictions on access to advanced computing hardware.

“Despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models.”

The release puts additional pressure on Chinese AI firms to defend their positions in the open-model market. Lower-cost systems from China are already attracting companies seeking alternatives to leading US platforms, particularly when pricing and customization matter as much as benchmark performance.

Open access brings infrastructure tradeoffs

For organizations across China and the wider APAC region, Kimi K3 could provide more control over customization, data handling, and local-language applications. Banks, telecommunications providers, government agencies, and software companies may be especially interested in an open-weight model that gives them greater oversight of deployment and adaptation.

That flexibility comes with a steep infrastructure burden. Running a 2.8-trillion-parameter model locally could require hundreds of thousands of dollars in computing equipment, according to an estimate cited by Reuters. Full on-premises deployment would therefore be beyond the reach of many organizations.

Cloud access may be the more practical option, reducing upfront hardware costs while introducing tradeoffs around data residency, vendor dependence, latency, and recurring usage fees. Businesses will also need to compare Kimi K3's accuracy, security controls, licensing terms, energy use, and total cost with smaller Chinese models and proprietary US alternatives.

Kimi K3 strengthens China’s position in the global open-model market. Its long-term impact will depend on whether organizations can turn its scale into reliable performance without allowing infrastructure costs to erase the advantages of open access.