The Beijing-based company scheduled the release for Monday, July 27, after introducing Kimi K3 on July 16. Opening the model weights will allow researchers and companies to run, examine and adapt the system on their own infrastructure rather than relying solely on Moonshot’s online service or application programming interface.
Kimi K3 has 2.8 trillion parameters, a one-million-token context window and native multimodal capabilities. It is designed to work with text, images and other forms of information while handling lengthy coding, research and knowledge-work assignments.
The scale places K3 in a category above Moonshot’s Kimi K2, which was released with one trillion total parameters in July 2025. Both use a mixture-of-experts architecture, where only a portion of the model is activated for each task. This approach can deliver the capacity of a very large system without using every parameter during each response.
Public access to the weights could extend K3’s reach among start-ups, universities and software developers seeking alternatives to proprietary models operated by OpenAI, Anthropic and Google. Users with sufficient computing resources can deploy the system locally, fine-tune it for specialised purposes and alter its behaviour.
The release also strengthens China’s position in open-weight AI. DeepSeek, Alibaba’s Qwen division, Zhipu AI and Moonshot have used downloadable models to attract international developers, particularly in markets where businesses are sensitive to the cost of proprietary services.
Moonshot has priced K3’s application programming interface at $3 per million input tokens and $15 per million output tokens, while cached input costs $0.30 per million tokens. Competitive pricing has become an important part of the challenge posed by China-based laboratories to larger US rivals, whose advanced systems often require higher expenditure for heavy commercial use.
K3 has drawn attention for its performance in coding, agent-based workflows, front-end development and complex knowledge tasks. Moonshot says the model competes with leading closed systems on several tests and can coordinate multiple agents working in parallel. Independent evaluations have also placed it among the strongest available models, although benchmark results vary depending on prompts, tools and testing conditions.
The company temporarily restricted some new paid subscriptions after demand created pressure on its computing capacity. Making the weights downloadable could shift part of that workload to users operating their own servers, although the model’s size means full deployment will require extensive hardware, memory and technical expertise.
The announcement has intensified debate in Washington over the security and economic implications of capable China-developed open models. US officials and technology executives are divided between those seeking tighter controls and those warning that broad restrictions could weaken open development within the United States.
Concerns include the possible use of downloadable models for cyber operations, surveillance and automated disinformation. Open weights can be modified after release, making central enforcement difficult. Supporters argue that transparent models can also be inspected for weaknesses, operated without sending sensitive information to foreign servers and adapted to meet local security requirements.
Moonshot has separately faced allegations that it used outputs from proprietary US systems to improve its technology through unauthorised distillation. US officials have raised the possibility of penalties, while the company’s critics have questioned whether advanced chips covered by export controls contributed to K3’s development. No publicly disclosed evidence has established that K3 itself was produced through unlawful copying, and Moonshot has presented the model as the result of its own research.
The dispute is unfolding as US technology groups debate whether open-weight models should be treated primarily as security risks or strategic assets. Nvidia, Microsoft, Meta and several other companies have supported preserving access to open AI development, while Anthropic has adopted a more cautious position on powerful downloadable systems.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
