A molecular-generation foundation-model family pretrained from scratch on 1.5B molecules, achieving state of the art on both de-novo and goal-directed generation. Released as multiple HuggingFace checkpoints, from Sarath Chandar's lab. NovoMolGen studies how molecular-language-model pretraining scales and transfers to downstream drug-discovery objectives. (Per-checkpoint parameter counts are not individually disclosed.)

Paper

sciencechemistrygenerativeopen-weight

Related