Moonshot AI Releases Kimi K3: World's First 2.8T Parameter Open Source Model Shocks the Industry

Kimi K3's technical architecture represents a significant breakthrough in large model design. The model adopts two innovative technologies: Kimi Delta Attention (KDA) and Attention Residuals, both architectural updates designed to improve how information flows across sequence length and model depth. Meanwhile, Moonshot AI significantly scaled up Mixture of Experts (MoE) sparsity, effectively activating 16 out of 896 experts when paired with a Stable LatentMoE framework. These structural improvements, combined with refined training and data recipes, yield an approximate 2.5× improvement in overall scaling efficiency compared to the previous generation Kimi K2, allowing the model to convert compute into intelligence more effectively.

In terms of performance, Kimi K3 demonstrates exceptional capabilities across multiple key benchmarks. In coding tasks, the model can handle complex long-horizon programming tasks, showing deep understanding of code logic and project structure. For knowledge work, Kimi K3's 1-million-token context window enables it to process ultra-long documents, large codebases, and complex research projects. In reasoning tasks, the model exhibits strong logical reasoning and problem-solving abilities. Notably, while Kimi K3's overall performance still slightly trails Claude Fable 5 and GPT 5.6 Sol — the two most powerful proprietary models — it consistently outperforms all other tested models, proving that open source models are fully capable of reaching frontier levels.

Kimi K3's release strategy is also noteworthy. The model is now fully available on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. At launch, Kimi K3 will use max thinking effort by default, with low- and high-effort modes to be introduced in subsequent updates. Moonshot AI stated they are working closely with inference partners and open-source maintainers to align technical details and ensure a reliable rollout across the ecosystem. The full model weights will be released by July 27, 2026, along with a detailed technical report on architecture, training, and evaluations. This phased release strategy ensures model stability while providing the community with adequate preparation time.

Kimi K3's release has profound implications for the entire AI industry. First, it proves that open source models can achieve performance levels comparable to proprietary models, which will accelerate the democratization of AI technology. Second, the 2.8T parameter scale demonstrates Moonshot AI's technical strength in large-scale model training, earning an important position for Chinese AI companies in global competition. Third, Kimi K3's open source strategy will promote the development of the entire ecosystem, allowing more developers and researchers to innovate based on this powerful foundation model. Finally, Kimi K3's success also shows that through innovative architecture design and training methods, frontier performance can be achieved without relying on the most advanced hardware, which is particularly important for research institutions with limited resources.