Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

Cactus Hybrid is a new model that can detect when it's wrong, reducing the need for expensive frontier models. It uses a confidence score between 0 and 1, allowing developers to route queries to a bigger cloud model when it's low. The model matches Gemini 3.1 Flash-Lite on most benchmarks and has been published on HuggingFace with open-source code. Developers can use it with various frameworks, including Transformers and Llama.cpp. The model has some caveats, including limitations on single-sequence decoding and hierarchical routing.