Jeremy Wohlwend

Jeremy Wohlwend is the cofounder and CTO of Boltz, where he builds molecular design models and tools. He is the first-listed author of the 2024 Boltz-1 report and a coauthor of the 2025 Boltz-2 report, contributing to a collaborative effort to model biomolecular interactions and predict binding strength. He completed his PhD at MIT under Regina Barzilay and uses the developer handle jwohlwend.[2][4][3]

Wohlwend's Boltz publications address two related questions: what a molecular interaction looks like, and how tightly the molecules bind. The team describes Boltz-1 as an open-source biomolecular interaction model approaching AlphaFold3 accuracy. Boltz-2 jointly models complex structures and binding affinities, extending the work toward molecular design. Gabriele Corso and Saro Passaro are coauthors of both reports, alongside larger research teams.[4]

The official Boltz repository distinguishes two affinity outputs with different uses. A predicted binder probability helps separate potential binders from decoys during hit discovery; a separate affinity estimate supports comparing binders and assessing small molecular modifications during optimization. The repository provides code and model weights under the MIT license for academic and commercial use, while marking updated Boltz-2 training and evaluation code as forthcoming.[4]

In their January 8, 2026 manifesto, Wohlwend, Corso and Passaro explained why they founded Boltz as a public-benefit corporation. Their effort began at MIT, but seeing scientists use the models convinced them that academic code releases needed an organization able to build, maintain and support usable products. Their stated ambition was: “We believe in a future where every scientist can iterate at the speed of inference.”[5]

The founders chose a software business: they planned to earn revenue by building and operating tools, and committed to avoiding drug development because they regarded competing with their users as contrary to their mission. On January 8 they launched Boltz Lab, with agents for small-molecule discovery and protein design, to address practical barriers including compute at scale, reliable infrastructure and interfaces suited to scientific workflows. They also committed to continued open-source models and open validation with academic and industry collaborators.[5]

Before the Boltz publications, Wohlwend worked on reducing the cost of language models. At EMNLP in November 2020, he published Structured Pruning of Large Language Models with Ziheng Wang and Tao Lei. The paper lists all three as equal contributors and gives each an ASAPP affiliation, documenting an earlier phase of his AI research in model compression.[6][7]

Their method represents weight matrices through low-rank factorization and adaptively removes rank-1 components during training. After pruning, the remaining matrices are smaller but still dense: inference uses ordinary matrix multiplication and avoids storing the indices required for sparse weights. The authors applied the approach to language modeling, adaptive word embeddings and BERT fine-tuning benchmarks.[6][7]