Open source for the win!! On July 11th, Kimi K2 (@Kimi_Moonshot) launched after 6 months of closed-door building. The model performed outstandingly, and quickly claimed the title of “the best open-source LLM model”. In fact, the day after Moonshot announced Kimi K2, OpenAI CEO Sam Altman urgently postponed the release of their next open-source model due to “security reasons”, which left people wondering. Here’s a technical breakdown of the new state of the art LLM that not only rivals proprietary industry giants like ChatGPT and Claude, but also challenges the limits of a good open-source model👇🧵

2/5: Kimi K2 is a mixture-of-experts (MoE) model with 32 billion activated parameters trained on 15.5T tokens, with a knowledge base up to April of 2025. Kimi specializes in coding and agentic tasks, with a SWE-bench score surpassing all open-source peers, such as DeepSeek V3 and Qwen3. It even surpasses some proprietary models like OpenAI’s GPT-4.1 and Claude 4 Opus at single-prompt competitive coding tasks. It also excels at tool use, with one of the best open source Tau2 benchmark score, closely tailing Claude 4 Opus.

3/5: Kimi’s API is supported by OpenAI and Anthropic which means you can directly export the auth keys from and use them in Claude code to replace Opus or Sonnet. It also supports Cline, which means you can try Kimi K2 in most coding IDEs from its API. However, one slight thing to note is that, a minimum deposit of around $15 (100RMB) is required to unlock a reasonable RPM (500). The reason given by Moonshot is to prevent malicious attacks. While the caution is understandable given the model's popularity, future development should prioritize improving accessibility for users.

5/5: Want to learn more? Read their docs !