“Merging "thinking" and "non-thinking" capabilities into a single model, as done in DeepSeek V3.1, is an extremely difficult training challenge.”