MIT and Recursion needed an open model that could co-fold protein complexes and predict binding affinity with the speed of foundation-model inference instead of the cost and delay of physics-based simulation, but training that model required supercomputer-scale infrastructure and efficient GPU kernels.
Reported impact
Boltz-2 matched much slower physics-based approaches closely enough for practical discovery work, accelerated triangle operations by up to 5x, sped up NIM inference by 2x to 3x over the open implementation, and brought open-source protein co-folding and affinity prediction to the broader life-sciences community.
MIT's Jameel Clinic and CSAIL researchers partnered with Recursion to build Boltz-2, an open-source biomolecular foundation model trained on Recursion's BioHive-2 supercomputer. The team used an NVIDIA DGX SuperPOD composed of 63 DGX H100 systems tied together with NVIDIA Quantum-2 InfiniBand networking so the model could train on roughly 3 million assay-labeled examples and process the large structural workloads needed for protein co-folding. NVIDIA cuEquivariance kernels accelerated triangle operations and reduced memory requirements during training, while an accelerated NVIDIA BioNeMo NIM microservice packaged the model for production inference. The result is a model that predicts both structure and binding affinity in one workflow, can return results in about 20 seconds on a single A100 GPU, and can be retrained or fine-tuned by the broader research community because code and weights are released under an MIT license.
Assay-labeled biomolecular training data flows into the BioHive-2 DGX SuperPOD, cuEquivariance accelerates the co-folding model kernels during training, and the resulting weights can be served through the BioNeMo NIM microservice for structure-and-affinity inference workloads.
Boltz-2 NIM accelerates protein structure prediction inference up to 2x–3x over the open-source implementation.
Leveraging the BioHive GPU supercomputer, cuEquivariance kernels, and a ready-to-deploy NIM microservice, the team trained and released an open-sourced Boltz-2.
Training of the Boltz-2 model was enabled by Recursion's BioHive-2 Supercomputer.
cuEquivariance implementations accelerate triangle operations up to 5x.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Sep 22, 2026 DELFI Diagnostics DELFI Diagnostics needed to speed genomic analysis for cancer screening and therapeutic monitoring, but CPU-bound alignment and somatic variant-calling steps created turnaround-time and cost bottlenecks that limited how… 10–20 business days DeepSomatic · MutectCaller · NVIDIA Parabricks NVIDIA Sample Jul 10, 2026 Astellas Pharma Inc. Astellas needed to shorten the slow handoff between experimental scientists and computational chemists in early drug discovery so more researchers could use structure prediction and affinity screening without waiting on… More than 50 regular users and over 2,000 total users within four months NVIDIA BioNeMo · NVIDIA Boltz-2 NIM · NVIDIA DGX H100/H200 NVIDIA Sample Apr 23, 2026 Biofy Hospitals needed pathogen identification and antibiotic-resistance analysis much faster than five-day culture and antibiogram workflows, but CPU-bound genomic processing and limited AI expertise made it hard to turn… 100,000+ synthetic DNA sequences generated by the biology model Abby · H100 Tensor Core GPUs · NVIDIA CUDA-accelerated libraries 