Kimi K3 challenges open-weight AI ceiling with vast memory
A Reasoning Model That Remembers a Whole Book
The latest Chinese-built large language model, named Kimi K3, has demonstrated capabilities that rival top US competitors on coding and spreadsheet tasks. Moonshot AI, the Beijing-based company behind the model, published tests showing it can match or outperform US models in these specific areas. For years, open-weight AI models lagged behind their proprietary counterparts, but the gap has now narrowed to its tightest point ever, according to Aaron Snoswell, an AI accountability researcher at the Queensland University of Technology Generative AI Lab in Brisbane. K3 stands out because of its enormous working memory: one million tokens of text, which allows it to remember the contents of thousands of lines of code or an entire book at once. This large memory reduces the likelihood of hallucination, the production of fabricated information, says Niusha Shafiabady, a computational intelligence researcher at Australian Catholic University in Sydney. The model has 2.8 trillion parameters, making it the largest open-weight model ever released, says Rahul Shome, a robotics and AI researcher at the Australian National University in Canberra. However, that size also means K3 cannot run on personal devices and will require significant institutional investment to operate, Shome adds.
China’s Strategy to Shape Global AI Governance
The release of Kimi K3 on 16 July came just before the 2026 Artificial Intelligence World Conference in Shanghai, where Chinese President Xi Jinping announced the formal start of a global alliance for AI regulation. Xi stated that all countries should take a people-centered approach and develop AI for positive and good purposes, framing AI as a driver for shared prosperity and common security. The timing is significant, says Mehwish Nasim, an AI researcher at the University of Western Australia in Perth, because K3 launched just weeks after Anthropic’s Claude Fable 5 and days after OpenAI’s GPT-5.6. Nasim interprets this as China signaling its ambition not only to build frontier AI systems but also to help shape the international AI ecosystem and its governance. Joel Pearson, a cognitive neuroscientist at the University of New South Wales in Sydney who studies how AI impacts people’s lives, calls the release a turning point, with people calling it a Sputnik moment. Three days after the model’s launch, Moonshot AI paused new sign-ups because demand pushed its system close to processing capacity limits. The model is open-weight, like its predecessor Kimi K2 and DeepSeek models, meaning its core components are publicly available and can be downloaded and modified by researchers for free, while training details remain private. It can also be accessed via an application programming interface at lower cost than proprietary models from OpenAI, Google, and Anthropic, which are closed-weight and cannot be modified.
Toby Walsh, a computer scientist at the University of New South Wales, observes that necessity here was the mother of invention, suggesting that the chip restrictions may have actually spurred innovation. The performance and capabilities of open-weight models have historically lagged behind US proprietary ones, but that gap is now the tightest it has ever been, according to Snoswell. Pearson notes that Chinese open-weight models are changing how investors and users perceive the value of US frontier models, especially given that Google, Anthropic, and OpenAI spend billions of dollars training their models compared with the much lower budgets of Chinese companies. Users could also lose access to US models if the government decides to impose restrictions, Pearson adds, pointing to a recent incident where Anthropic temporarily disabled its most advanced models, Fable 5 and Mythos 5, for all users after the US government directed it to suspend access for foreign nationals. Access to Fable 5 was restored this month, but Mythos 5 remains available only to specific US organizations. Moonshot AI says the weights for K3 will be released on 27 July, making it the largest open-weight model available, Shome confirms. The model is still too large for personal devices, however, meaning widespread adoption will require significant infrastructure investment. Moonshot AI has not disclosed a timeline for making the model accessible on smaller hardware
Sources
1. Moonshot AI
2. Australian Catholic University
3. Australian National University
4. University of Western Australia
5. University of New South Wales
6. Anthropic
7. Google
8. DeepSeek