Imagine an AI chemist, capable of dreaming up entirely new molecules with specific properties. This isn't science fiction, but the goal of a new area of AI research: text-based open molecule generation. Researchers have just introduced TOMG-Bench, the first benchmark designed to test how well large language models (LLMs) can tackle this complex task. It turns out that while LLMs excel at many text-based challenges, creating molecules from scratch is surprisingly difficult. TOMG-Bench tests LLMs on three core abilities: editing existing molecules, optimizing molecules for desired properties like drug-likeness, and generating completely novel molecules based on specific criteria. The results are a mixed bag. While leading proprietary models like Claude-3.5 show some promise, even they struggle to consistently generate valid, novel molecules. Interestingly, larger LLMs generally perform better, highlighting the importance of scale in AI. However, the benchmark also revealed that simply training LLMs on existing molecule-text datasets isn’t enough. The researchers found that models fine-tuned on these datasets often fall short on the open-ended tasks in TOMG-Bench. To address this, they created OpenMolIns, a new instruction-tuning dataset designed specifically for open molecule generation. When trained on OpenMolIns, even smaller LLMs showed significant improvement, suggesting that specialized training data is key to unlocking AI's potential in this field. The creation of TOMG-Bench and OpenMolIns is a major step forward. It provides a crucial testing ground for evaluating and improving AI-driven molecule design, opening doors to accelerating drug discovery, materials science, and more. The challenge now is to refine these models, develop even better training data, and push the boundaries of what AI can achieve in the molecular world.