1.7B model leading strict-7 formal reasoning above Qwen3-8B and Gemma-4-26B - specialists eating generalist territory?