Kilo has deployed a mysterious cutting-edge model from China's top laboratories, which nearly outperforms all publicly available weight models in long context encoding. If this truly comes from DeepSeek, it’s incredibly shocking. Codename: “Super Potato” 256k context window + 32k output limit + strict system prompt requirements, which are indeed very useful for production codebases 0xd488f93935bdfba32eb54559a9f7989dfd7e4444