Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong (github.com)

Cactus has introduced an innovative feature with its Gemma 4 E2B Hybrid model that allows AI to self-assess its responses for accuracy. This is achieved by embedding a confidence scoring probe within the model's architecture, which scores each answer on a scale from 0 to 1. If confidence falls below 0.85, the system can defer to a larger model for more accurate responses. The rollout begins with the Gemma 4 E2B Hybrid, designed to maintain high performance while utilizing minimal resources, effectively routing only 15-35% of queries to the more powerful Gemini 3.1 Flash-Lite model. This advancement holds significant implications for the AI/ML community, promoting both privacy and efficiency by enabling on-device decision-making. The model's architecture not only optimizes performance—matching the capabilities of its larger counterpart—but also improves response reliability by incorporating a method for determining when it is uncertain. The promising benchmarks reflect that the confidence-probing technique can make predictions across different modalities, suggesting the model's robustness in delivering accurate answers even when trained without specific audio data. As developers explore this technology, further optimization and benchmarking options open avenues for enhanced model utility in real-world AI applications.