Tech • IA • Crypto
Is This the Biggest AI Release of 2026? (China’s New DeepSeek Moment)
8/10TL;DR
China’s Moonshot AI has unveiled Kimi K3, a 2.8 trillion-parameter open-weight model nearing top U.S. systems, intensifying global competition over AI technology and governance.
KEY POINTS
Record-scale open model
Kimi K3 features 2.8 trillion parameters, making it the largest open-weight AI model announced to date. Full weights are expected to be released on July 27, allowing companies to download, modify, and run it independently. This marks a shift from closed ecosystems toward customizable, enterprise-controlled AI deployment.
Performance rivals top U.S. systems
Benchmarking places K3 close to leading proprietary models. It scored 57 on the Artificial Analysis index, roughly matching Gemini 3.1 Pro and slightly ahead of GPT 5.6 Terra and Claude Opus 4.8, while trailing top-tier models by only a few points. In coding tasks, it reportedly outperforms several major Western systems.
Top ranking in real-world coding tests
In Arena’s front-end coding benchmark, K3 ranked first with 1,679 points, beating Claude Fable 5 and GPT 5.6 Soul. It led in six of seven categories, including dashboards and branding tools, jumping from 18th place in the previous generation to the top in a single release cycle.
Designed for long autonomous workflows
K3 targets extended, complex tasks such as software development. It can analyze large codebases, plan changes, execute iterations, and verify results over hours or days. Its “vision in the loop” capability allows it to visually assess outputs and refine them, supporting applications like game development and UI design.
Massive context and multimodal capabilities
The model processes up to 1 million tokens (around 750,000 words) in a single session and handles text, images, and video together. Demonstrations include building a browser-based 3D game, simulating a Long March 10 rocket, and reproducing astrophysics research in hours rather than weeks.
Efficient architecture at extreme scale
K3 uses a mixture-of-experts design with 896 experts, activating only a subset per task to improve efficiency. New techniques such as Delta Attention and lower-precision training formats reduce compute demands, though deployment still requires large-scale infrastructure, with recommendations of 64 AI accelerators.
Aggressive pricing pressures competitors
Pricing is positioned below leading Western models: about $3 per million input tokens and $15 per million output tokens, compared with higher rates from rivals. For large enterprises already spending millions monthly on AI, a near-parity open model could significantly reduce costs while improving data control.
Market and geopolitical impact
The release triggered immediate market reactions, with Chinese AI stocks such as JIEPU and MiniMax falling sharply. It also echoes earlier disruptions from DeepSeek, which erased significant market value. Meanwhile, Chinese firms continue accelerating, with multiple models exceeding 1 trillion parameters.
China pushes alternative AI order
At the World Artificial Intelligence Conference in Shanghai, President Xi Jinping outlined a strategy to shape global AI governance. China is promoting open models, international cooperation, and initiatives like the World AI Cooperation Organization (WICO), which already includes 29 countries. The approach emphasizes affordability and broader access, especially across developing regions.
Rising tensions over standards and training
Disputes are growing over practices like model distillation, with U.S. firms alleging misuse of their systems. At the same time, debates continue over regulation, export controls, and data sourcing, highlighting a widening divide between U.S. and Chinese approaches to AI development and control.
CONCLUSION
Kimi K3 signals that open, large-scale AI from China is rapidly closing the gap with leading proprietary systems, while reinforcing a broader strategic push to reshape global AI markets, standards, and alliances.