Chinese AI company Moonshot AI officially released its latest large language model, “Kimi K2,” on July 12, 2025. The new model not only surpasses several Chinese open-source models including DeepSeek-V3 and Qwen3-235B in performance, but also outperforms commercial flagship models such as GPT-4.1 and Claude Sonnet 4 across multiple benchmark tests.
China’s Moonshot AI team has launched a next-generation open-source AI model, Kimi K2, touted as surpassing GPT-4.1.
Adopting a 1-trillion-parameter MoE architecture, purpose-built for AI agents.
Kimi K2 is a language model series primarily developed by Moonshot AI. The name “Kimi” continues from the previous generation model. This model adopts the Mixture of Experts (MoE) architecture, with a total of approximately1 trillion parameters，among which the active parameters (Active Parameters) are32 billionThis architecture allows the model to dynamically select subsets of sub-models to participate in the inference process, optimizing computational resource usage while maintaining model performance.
Unlike OpenAI’s o-series reasoning models, Kimi K2 is designed as an open agentic intelligence model. This means Kimi K2 is not just good at text-based Q&A; it can also integrate tools, complete task workflows, and even possess the ability to actually operate web pages, check emails, and plan itineraries, bringing it closer to true multimodal AI agent application scenarios.
Kimi K2’s release version includes two main branches:
-
Kimi-K2-BaseA base version without instruction fine-tuning, suitable for further custom development.
-
Kimi-K2-InstructA version that has undergone post-training (Instruction Tuning) to optimize the user interaction experience.
In multiple standardized language model benchmark tests, Kimi-K2-Instruct demonstrated impressive results. Not only inIt surpassed well-known open-source models such as DeepSeek V3 and Qwen3-235B across all test categories.and even outperformed commercial models such as GPT-4.1 and Claude Sonnet 4 in some tests, demonstrating its high performance in tasks like semantic understanding, logical reasoning, and knowledge retrieval.
AI Assistant Capability Showcase: Travel Planning and HTML Game Generation, All in One Go
Moonshot AI provided examples of Kimi K2’s real-world applications in its official release, demonstrating its powerful capabilities as an AI agent. One example showed that after Kimi K2 received a request to “help plan a trip,” it automatically completed the following actions:
-
Extract date information provided by the user.
-
Search the web for content related to attractions and weather.
-
Browse the user’s mailbox to confirm available times or tickets.
-
After consolidating the above information, output aVisual itinerary planner
Furthermore, on platform X, users have also reported that after Kimi K2 received the instruction ‘Please create a simple brick breaker game in a single HTML page,’Instantly produce a fully functional web game.Such cases show that Kimi K2 is also highly usable for code generation.
@Kimi_Moonshot k2 just oneshotted this game with the prompt “create a simple breakout game as a single html page”, cost less than a penny on @OpenRouterAI pic.twitter.com/DZRB1VZpYk
— waterdoggie (@waterdoggie) July 12, 2025
Hugging Face and GitHub are now fully open
The Kimi K2 development team has already been on the well-known machine learning platform.Hugging FaceReleased the complete data packages for the Kimi-K2-Base and Kimi-K2-Instruct models, making them available for researchers and developers to download and use. The model page includes architecture documentation, usage guides, training details, and performance data, making it easy for developers to quickly deploy or build upon the models. Those who don’t mind that the model was developed by a Chinese team can give it a try.
👉 Kimi-K2 – Moonshot AI’s model collection on Hugging Face
Meanwhile, Moonshot AI also open-sourced the complete source code for Kimi K2 on GitHub, including model invocation examples, deployment workflows, and API endpoints, supporting the community in continuously contributing to and optimizing the model.
👉 GitHub – MoonshotAI/Kimi-K2 Official Repository
Source: KOCPC Chinese