Moonshot AI has launched Kimi K3, the world's first open AI model with 2.8 trillion parameters and advanced capabilities.
Moonshot AI, a Chinese artificial intelligence startup, has launched Kimi K3, the first open AI model to feature 2.8 trillion parameters, a native vision system, and the ability to handle a context length of 1 million tokens. This new model is positioned as a strong competitor to proprietary models such as Claude Fable 5 and GPT 5.6 Sol.
Kimi K3 utilizes an innovative architecture known as Kimi Delta Attention (KDA) combined with Attention Residuals (AttnRes). This design enhances the model's ability to process lengthy data sequences efficiently while ensuring that important information is extracted from multiple layers, thereby improving contextual comprehension and retaining meaning in complex texts.
Efficiency is further boosted through the Stable LatentMoE framework, which operates only 16 out of 896 experts at once. This approach, along with updated training methods, has resulted in a significant increase in operational efficiency compared to its predecessor, K2.
For most of the past year, Kimi has maintained the largest size among open models, according to its developers. Kimi K3 is accessible through the Kimi Work platform, Code, and API, with the maximum reasoning mode set as default. Additional modes will be rolled out in the future, and a detailed report along with full weights will be available on July 27.
Performance Metrics
Benchmark tests indicate that Kimi K3 outperformed both Fable 5 and GPT 5.6 Sol in certain areas, including the SWE Marathon and BrowseComp, though it lagged behind in others like FrontierSWE and HLE-Full. The varied testing methods employed for these comparisons included Claude Code, Codex, and the proprietary KimiCode.
Capabilities in Programming and Automation
Kimi K3 is adept at conducting extensive engineering tasks with minimal human intervention, efficiently navigating large code repositories and managing terminal tools. In a notable test, it autonomously executed optimization tasks on GPU cores for 24 hours, achieving results comparable to Fable 5 while surpassing other models like Opus 4.8.
Additionally, K3 independently developed MiniTriton, a compact compiler designed for GPU cores, showcasing its advanced capabilities in engineering tasks.
Innovations in Chip Design and Game Development
Demonstrating its versatility, K3 autonomously designed a chip tailored for a neural network, completing the task in 48 hours using open-source tools. The chip, measuring just 4 mm², operates at 100 MHz and simulates over 8,700 tokens per second, featuring a sophisticated architecture with numerous components.
Beyond chip design, K3 can integrate 3D reasoning, coding, and vision to produce game prototypes based on concepts, images, and videos.
Research and Video Production
K3 has also shown prowess in research tasks, completing complex astrophysics relationships in a fraction of the time required by human researchers. It processed extensive scientific literature and generated thousands of lines of code, presenting findings interactively. In another instance, it compiled a comprehensive report on the ASIC chip industry after analyzing thousands of pages of documents.
Moreover, K3's video editing capabilities allow it to create animations and teasers autonomously, significantly reducing the time typically needed by human editors.
Challenges and Future Outlook
Despite its advancements, the K3 model has limitations, including sensitivity to losing reasoning history during environment transitions and a tendency to act excessively in unclear situations. Its overall usability currently trails behind that of Fable 5 and GPT 5.6 Sol, according to the development team.
Moonshot AI previously released Kimi K2 in July 2025, which featured 1 trillion parameters and quickly became a frontrunner in open systems for agent tasks.
