Moonshot AI Releases Kimi K3 Open-Source Model with 2.8 Trillion Parameters, Performance Close to Fable 5

On July 16, 2026, Moonshot AI (Beijing) released the Kimi K3 model, with 2.8 trillion parameters and a 1 million token context window, making it the world's largest open-source weights model.

Fact Check

The model natively supports visual understanding and is optimized for software engineering, knowledge work, deep research, and multimodal understanding scenarios. According to official information, Kimi K3 is scheduled to open its model weights on July 27, and currently provides API access mainly through the Kimi AI platform. Reuters reported that the United States previously suspended access to Anthropic's Mythos and Fable models for foreign users, creating a contrast with the release of Kimi K3.

Mechanism Breakdown

Kimi K3 adopts a MoE architecture with 2.8 trillion total parameters, activating 16 out of 896 experts during actual inference. During training, it uses self-developed Kimi Delta Attention and Attention Residuals mechanisms, achieving approximately 25% improvement in training efficiency at an additional cost of less than 2%, along with up to 6.3x decoding acceleration. These structural improvements result in an overall scaling efficiency improvement of about 2.5x compared to Kimi K2. Bloomberg reported that Moonshot AI is seeking a valuation of around $30 billion in its latest funding negotiations, a leap from the previous several billion dollars.

Industry Impact

In terms of competitive landscape, the release of Kimi K3 highlights the narrowing technology gap between China's open-source models and the US's frontier AI systems. Xinhua News Agency reported that Chinese AI companies, including Moonshot AI, Zhipu, and MiniMax, are rolling out more powerful models at lower costs with shorter release cycles. For developers, Kimi K3 provides an open-weight path, allowing local deployment and modification, reducing reliance on closed-source APIs. For enterprise users, the model approaches Fable 5 in GPU kernel optimization tests, making it suitable for handling long-duration programming and multimodal tasks.

Comparison and Precedents

A Reuters report noted that the US suspended foreign users' access to Anthropic's models due to national security considerations, directly contrasting with Kimi K3's open-source path. Historical data shows that Chinese AI companies are achieving efficient scaling under resource constraints through MoE architecture, differing from the previous paradigm dominated by closed-source models.

Strategic Assessment

Based on existing facts, Kimi K3 is most likely to drive more developers toward the open-source path.

In terms of selection recommendations, developers who need to handle million-token context and vision tasks can prioritize testing the Kimi K3 API; enterprises that value long-term controllability should pay attention to local deployment options after weight release, while comparing with Fable 5's closed-source limitations. Moonshot AI has received support from Alibaba and Tencent, and the valuation leap at the capital level reflects market recognition of the open-source path.

In practical applications, Kimi K3 achieves a visual closed loop in code and real-time screenshot iteration, making it suitable for game development and digital creation scenarios. Multiple third-party evaluations show its overall performance ranks among the global first tier, yet gaps remain compared to the first and second positions globally. Enterprises can first validate via API, then decide whether to wait for the weight release.