Boltz

Boltz is a public-benefit AI research company building models and software for biomolecular structure prediction, binding affinity estimation and molecular design. Its MIT research origins produced Boltz-1, Boltz-2 and BoltzGen; the company subsequently launched Boltz Lab and proprietary design pipelines. Gabriele Corso and Jeremy Wohlwend are co-founders, and Corso is CEO.[8][7][1][12][10]

Boltz publicly announced its company launch on January 8, 2026. In its manifesto, Gabriele Corso, Jeremy Wohlwend and Saro Passaro explained why the MIT effort needed an organization beyond academic code releases: “It needs products, and an organization that can build, maintain, and support them over time.” Boltz PBC would combine research, products and community while continuing to release open models.[7][8]

The founders committed to making money by building and operating software rather than developing drugs. The initial business model charged for usage, earning a margin over optimized compute costs, with additional services for model customization and enterprise deployment. Boltz also committed to customer ownership of outputs and to not training on customer data.[8]

At launch, Boltz announced a $28 million seed round led by Amplify, a16z and Zetta, with additional backing including Clement Delangue, Factorial Capital and Obvious Ventures. It reported that its model suite had been used by more than 100,000 scientists across academia and industry. That figure describes company-reported model adoption, rather than paying customers.[7][8]

Boltz Lab launched in beta with protein-design and small-molecule-design agents. These workflows address tasks beyond a single prediction: preparing targets, generating candidates, ranking designs and selecting experiments. As described in the February 12, 2026 interview, the agents automate pipelines of models; the small-molecule workflow used a generative model that learned appropriate building blocks to search a synthesizable chemical space.[7][1]

Boltz Lab combined parallel GPU infrastructure with an API and a collaborative interface where scientists could rank candidates and build consensus about which molecules to test. Its small-molecule screening pipeline ran ten times faster than the open-source version, according to the February 2026 interview. The claim concerned that pipeline, rather than every Boltz model or scientific workflow.[1]

The models underlying the company preceded its January 2026 public launch. Earlier academic work also supplied context for its emphasis on testing generalization: according to the February 2026 interview, feedback on protein–small-molecule interaction predictions exposed failures on proteins unlike those used in training. Working with Nick Polizzi’s Harvard group, the researchers developed a harder benchmark before pursuing improvements.[4][5][6][7][1]

Boltz-1, released on November 16, 2024, predicted the three-dimensional structures of biomolecular complexes. Jeremy Wohlwend, Gabriele Corso and Saro Passaro announced it on behalf of a broader academic team including Tommi Jaakkola and Regina Barzilay. The release placed training and inference code, weights and training data under the MIT license, permitting academic and commercial use. The team credited MIT Jameel Clinic, Genesis Therapeutics for engineering, infrastructure and compute, and the U.S. Department of Energy for computational support.[4]

Training Boltz-1 involved severe compute constraints. According to recollections in the February 2026 interview, the researchers had enough compute to train the large model only once, stopping and restarting the run to fix bugs without returning to the beginning. Training on a shared Department of Energy cluster proceeded in two-day windows followed by week-long queue waits, with Genesis helping finish the model.[1]

Boltz-2, announced on June 6, 2025, expanded from predicting interaction geometry to estimating how tightly a protein and ligand bind. Developed by the MIT Jameel Clinic team alongside Recursion, it added an affinity module, synthetic and molecular-dynamics training data, GPU optimizations and controls such as templates and contact constraints. Boltz reported an average Pearson correlation of 0.62 on the held-out-target FEP+ benchmark, comparable to OpenFE while running more than 1,000 times faster in that comparison.[5]

The Boltz-2 announcement reported that thousands of scientists, including teams at all twenty largest pharmaceutical companies, had used Boltz-1. Boltz-2 retained MIT licensing for academic and commercial use. The official repository distinguishes binder-versus-decoy probability from an affinity estimate intended for ligand optimization; as of September 18, 2026, it still listed updated Boltz-2 training and evaluation code as forthcoming.[5][15]

BoltzGen, announced by Hannes Stark on October 26, 2025, moved into generating new protein and peptide binders. Developed at MIT, it unified design and structure prediction in one all-atom generative diffusion model. A geometry-based representation encoded designed residue types, while a specification language controlled binding sites, covalent bonds, secondary structures and design masks. Downstream validation, redesign and ranking narrowed the generated pool to a small, diverse set for experimental screening.[6]

BoltzGen’s experimental validation involved multiple academic and industry laboratories. On a panel of nine targets with no known binders and less than 30% sequence similarity to bound molecules or complexes in the Protein Data Bank, the team reported nanomolar nanobody binders for six targets, testing fifteen designs or fewer per binder-target pair. BoltzGen was also released under the MIT license; the binding results established experimental starting points rather than clinical efficacy.[6][1]

On June 16, 2026, Boltz announced BoltzMol-1, BoltzProt-1 and the Boltz API. BoltzMol-1 used a version of Boltz-2 optimized for prospective screening, either ranking purchasable compounds or generating and searching a make-on-demand space exceeding 74 billion compounds. Boltz reported confirmed hits on six of ten challenging targets after testing 28–51 compounds per target. Collaborations included Anthony Gitter and Nathan Wlodarchak on tuberculosis kinase PknB, and the Kritzer Lab at Tufts on autophagy proteins LC3B and GABARAP.[10]

BoltzProt-1 improved both generation and scoring. Its Boltz-PPI interaction model, trained on structural and patent-derived data, scored proposed protein interactions more directly than structural confidence alone. Boltz reported nearly triple BoltzGen’s nanobody hit rate across ten hard benchmark targets and 58% of generated binders passing a complete multi-criteria developability panel. The company stressed that these designs still required downstream optimization.[10]

The API served BoltzMol-1, BoltzProt-1 and Boltz-2, with Python and JavaScript SDKs and support for Claude Code, Codex and Gemini CLI. Its advertised launch price started at $0.025 per prediction. Distribution partners included Benchling, Phylo, Amazon Bio Discovery, Rowan, Tamarind, Kiin Bio, Pauling.ai, Mirror Physics and Cultivarium. Boltz credited NVIDIA’s cuEquivariance kernels for efficient inference.[10]

A September 15 case study by Geoffrey Smith illustrated BoltzMol-1’s screening workflow on Werner helicase. Boltz screened 171,751 purchasable WuXi compounds, selected 27 for experiments and reported three functional IC50 values—8.4, 14.4 and 27.9 µM—with two compounds also confirming binding by surface plasmon resonance. The reported total cost was about $10,000, including $4,000 in compute. A single chemist completed two days of computational work followed by a week of assays at a CRO partner.[14]

The WRN campaign used the known HRO761-bound structure, PDB 8PFO, as an inference template, although that structure postdated model training. Earlier WRN structures and 883 recorded WRN inhibitors were present in training-related data. Smith acknowledged hindsight and described the results as putative hits requiring further confirmation, including structure–activity relationships, orthogonal biophysics and ADME profiling.[14]

Boltz paired open research releases with proprietary enterprise models. Its January 8 Pfizer collaboration, announced as multi-year, called for refining models on Pfizer’s historical data and building custom workflows for selected programs. The June 18 Takeda agreement covered Boltz Lab, API and proprietary model access across Takeda Research, plus target-specific fine-tuning. The July 2 GSK collaboration similarly planned exclusive models and workflows adapted to GSK data. Each agreement left ownership of resulting compounds with the pharmaceutical partner.[7][9][11][12]

On September 14, 2026, Boltz announced a partnership with dsm-firmenich extending its work into scent and flavor ingredient discovery. The planned collaboration combined Boltz’s proprietary models with taste and odor receptor data to create exclusive prediction and design models, with platform access and embedded Boltz researchers working alongside dsm-firmenich teams.[13]