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

Cactus Compute has developed Cactus Hybrid, a technique that trains small language models like Gemma 4 E2B to assess their own confidence in answers with a 0-1 score, enabling intelligent routing to larger models only when confidence falls below a threshold. Using this approach, Gemma 4 E2B matches Gemini 3.1 Flash-Lite's performance on most benchmarks while handling 65-85% of queries on-device and routing only 15-35% to the larger model, achieving privacy and speed benefits with minimal quality loss.