NVIDIA BioNeMo

NVIDIA BioNeMo is NVIDIA's software platform for applying AI to biology and drug discovery. It began in 2022 as a hosted cloud service for biomolecular language models, was reorganised in November 2024 around an open-source training framework plus packaged inference microservices and reference workflows, and has since grown into a collection of public model releases, GPU-accelerated libraries, agent skills and, from September 2026, a dedicated inference runtime for structure prediction. ^[1][2]

The name is a deliberate echo of NVIDIA NeMo, the company's framework for general-purpose language and agent models, and the two were announced together. They are separate product lines: NeMo addresses text, speech and agents, while BioNeMo addresses proteins, nucleic acids, small molecules and single-cell data. ^[1]

Infobox

Origins as a cloud service

NVIDIA announced BioNeMo on 20 September 2022 at GTC, in the same press release as the NeMo Large Language Model Service. The BioNeMo LLM Service was described as a cloud API that extended large language model techniques beyond text into scientific applications, with support for protein, DNA and biochemical data, and it shipped with two new BioNeMo language models for chemistry and biology. NVIDIA's framing at the time was that models with billions of parameters could store more information about protein structure and evolutionary relationships between genes, and could generate novel biomolecules. ^[1]

That positioning placed BioNeMo in the same family as protein language models such as ESM, but as managed infrastructure rather than a published model. The service-first approach lasted about two years.

The 2024 reorganisation

On 18 November 2024, at SC24, NVIDIA announced that the BioNeMo Framework was open source and described a three-part platform. ^[2]

The same release named the first NIM microservices for BioNeMo and attached performance claims to two of them. NVIDIA said researchers using the AlphaFold2 microservice achieved a fivefold speedup in protein structure prediction, and that the DiffDock 2.0 microservice, trained on the PLINDER dataset, predicted molecular orientation 6.2 times faster and 16 percent more accurately. RFdiffusion and ProteinMPNN microservices covered protein design and inverse folding. The release also introduced cuEquivariance, an acceleration library for the equivariant mathematics that DiffDock and related chemistry models depend on. All of these figures are NVIDIA's own. ^[2]

NVIDIA listed contributors to the open-source framework including A Alpha Bio, Argonne National Laboratory, Dyno Therapeutics, Genentech, Ginkgo Bioworks, Relation, VantAI and Weights and Biases, and said more than 200 techbio companies, large pharmaceutical firms and startups were integrating BioNeMo into their computer-aided drug discovery platforms. Kimberly Powell, NVIDIA's vice president of healthcare, framed the release around the convergence of AI, accelerated computing and growing biological datasets. ^[2]

Framework and Recipes

The BioNeMo Framework is a set of Python sub-packages built on NVIDIA's NeMo 2 and Megatron stack, covering data handling, model definitions and training loops for biological foundation models. Its documented sub-packages include bionemo-esm2, bionemo-geneformer, bionemo-evo2, bionemo-amplify, bionemo-llm, bionemo-moco for generative interpolants, bionemo-scdl for single-cell data loading, bionemo-noodles for sequence file access, and bionemo-size-aware-batching. ^[7]

The repository that housed the framework now carries the name BioNeMo Recipes, and a request for the old NVIDIA/bionemo-framework path resolves to NVIDIA-BioNeMo/bionemo-recipes, created in October 2023. ^[3][5] The stated design change is a move away from a batteries-included training framework toward two narrower products: Hugging Face compatible model checkpoints whose layers use NVIDIA TransformerEngine, and standalone training recipes that demonstrate how to reach high throughput with common frameworks. ^[3]

Recipes ship for native PyTorch, Hugging Face Accelerate and PyTorch Lightning, and the models covered include ESM-2, AMPLIFY, Geneformer, CodonFM, a vision transformer, and several general-purpose architectures used as scaling references: Llama 3, a Mixtral-style mixture-of-experts model, and Qwen 2.5 and 3. The published support grid tracks which recipes handle fully sharded data parallel training under FSDP2 and megatron-FSDP, BF16, FP8, MXFP8, NVFP4, sequence packing, and context parallelism, and marks unfinished combinations rather than claiming them. ^[3] The Recipes repository is licensed under the Apache License 2.0. ^[3]

NIM microservices

BioNeMo's inference side is packaged as NVIDIA NIM microservices: containers that bundle a model with its runtime optimizations behind an API endpoint. The catalogue published with the BioNeMo Agent Toolkit in 2026 lists NIM microservices for Boltz-2, DiffDock, Evo 2, GenMol, MolMIM, MSA-Search (which builds alignments through ColabFold), OpenFold 2, OpenFold3, ProteinMPNN and RFdiffusion. ^[4]

Two composite workflows are published alongside the individual services: a generative virtual screening chain that runs GenMol into DiffDock into Boltz-2, and an MSA-enabled structure prediction chain that runs MSA-Search into OpenFold3. ^[4] These correspond to the Blueprint concept introduced in 2024. ^[2]

BioNeMo Agent Toolkit

The BioNeMo Agent Toolkit, whose public repository dates from June 2026, packages NVIDIA's life-science tools as skills that a coding or scientific agent can call. Each skill supplies instructions, scripts and references for selecting a tool, preparing its inputs, running it and interpreting the output. The catalogue spans the NIM microservices above, open model releases such as Proteina-Complexa and KERMT, and libraries including Parabricks for genomics, nvMolKit for cheminformatics and cuEquivariance. Skills install through a command-line tool and through plugin marketplaces for individual agent harnesses. ^[4][5]

BioNeMo Inference Runtime

NVIDIA BioNeMo Inference Runtime (BioIR) entered public beta on 10 September 2026. It is a PyTorch library that accelerates biomolecular structure prediction on NVIDIA GPUs using fused kernels, CUDA Graph capture and a Ray based pipeline that places one full model replica on each GPU. It runs AlphaFold2 and OpenFold2 variants, OpenFold3, Boltz-1 and Boltz-2 through an end-to-end processor, exposes Protenix v2 and a Boltz-2 affinity model as modules, and lets developers drop individual optimized modules into architectures of their own. The code is Apache-licensed, though the CuTeDSL kernel sources are not published and ship as precompiled binaries. ^[6][8]

BioIR occupies a different position in the stack from the NIM microservices. The microservices are hosted or self-hosted endpoints for people who want predictions; BioIR is a library for people building or scaling the models themselves, and NVIDIA's launch post for BioIR points readers who want agentic orchestration at the Agent Toolkit rather than at the runtime. ^[8]

Open model and library releases

NVIDIA publishes a substantial part of its digital-biology research through the NVIDIA-BioNeMo GitHub organisation. The repositories, with NVIDIA's own descriptions, include:

Proteome-scale structure prediction

NVIDIA's largest published result in this area is its part in the expansion of the AlphaFold Protein Structure Database to protein complexes, a collaboration with EMBL's European Bioinformatics Institute, Google DeepMind and Seoul National University announced on 16 March 2026. The accompanying preprint reports predictions for 23,441,822 homodimers and 7,620,644 heterodimers across 4,777 proteomes, of which 1,754,242 high-confidence homodimers were released through the database. NVIDIA's contribution was described by EMBL-EBI as the methodology work alongside the Steinegger Lab, including accelerations to multiple sequence alignment and deep learning inference, plus the infrastructure and scaled-out inference pipelines; the preprint describes the accelerated inference path as an OpenFold implementation using TensorRT and cuEquivariance, run on a DGX H100 SuperPOD. EMBL-EBI put the compute the collaboration absorbed on the community's behalf at around 17 million GPU hours. ^[9][10]

See also

• NVIDIA NIM
• AlphaFold
• Protein folding
• Evo 2
• NVIDIA DGX

References

1. ^1 ^2 ^3 ^4NVIDIA. "NVIDIA Launches Large Language Model Cloud Services to Advance AI and Digital Biology", NVIDIA Newsroom, 20 September 2022. nvidianews.nvidia.com/...ce-ai-and-digital-biology
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11NVIDIA. "NVIDIA Opens BioNeMo to Scale Digital Biology for Global Biopharma and Scientific Industry", NVIDIA Newsroom, 18 November 2024. nvidianews.nvidia.com/...a-and-scientific-industry
3. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA. "BioNeMo Recipes" README and LICENSE, GitHub, retrieved 11 September 2026. github.com/...bionemo-recipes
4. ^1 ^2 ^3 ^4NVIDIA. "NVIDIA BioNeMo Agent Toolkit" README, GitHub, retrieved 11 September 2026. github.com/...bionemo-agent-toolkit
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14NVIDIA BioNeMo organisation repository listing, GitHub REST API, retrieved 11 September 2026. github.com/NVIDIA-BioNeMo
6. ^1 ^2NVIDIA. LICENSE (Apache License 2.0), BioNeMo-Inference-Runtime repository, retrieved 11 September 2026. github.com/...LICENSE
7. ^NVIDIA. BioNeMo Framework documentation, retrieved 11 September 2026. docs.nvidia.com/...latest
8. ^1 ^2Tretina, Kyle; Tal, Roy; Salehi, Mahan; Patel, Neel. "High-Throughput Structure Prediction with BioNeMo Inference Runtime", NVIDIA Technical Blog, 10 September 2026. developer.nvidia.com/...-bionemo-inference-runtime
9. ^Stroe, Oana. "Millions of protein complexes added to AlphaFold Database shed light on how proteins interact", EMBL-EBI news, 16 March 2026, with update of 19 May 2026. ebi.ac.uk/...first-complexes-alphafold-database
10. ^Han, Yewon; Tsenkov, Maxim I.; Venanzi, Niccolo A. E.; et al. "AlphaFold Database expands to proteome-scale quaternary structures", preprint, NVIDIA Research. research.nvidia.com/...afdb.pdf