NVIDIA BioNeMo
BioNeMo is NVIDIA's development platform for AI in biology and drug discovery: open models, training recipes, libraries, NIM microservices and, since June 2026, an agent toolkit. The older BioNeMo Framework container is archived; NVIDIA now points users to BioNeMo Recipes on GitHub.123
Also known as BioNeMo Framework, BioNeMo Recipes, BioNeMo Agent Toolkit
At a glance
- What is it?
- BioNeMo is a collection of open resources rather than one installable product. It has four main parts: BioNeMo Recipes, which provides model code and training recipes for biological foundation models such as ESM-2, Geneformer, AMPLIFY and CodonFM; research models published in the NVIDIA-BioNeMo GitHub organization, whose index also points to GPU libraries hosted there or in other repositories; NIM microservices that serve models such as Boltz-2, OpenFold and Evo 2 behind an API; and the BioNeMo Agent Toolkit, a set of skills that lets coding and research agents run these tools. The former BioNeMo Framework container on NGC is archived and no longer maintained.1456
- What does it do?
- BioNeMo helps teams train, adapt and run AI models for proteins, DNA and RNA, single cells and small molecules on NVIDIA GPUs. Recipes show how to train transformer models efficiently at scale with TransformerEngine, FSDP, FP8 and context parallelism. Hosted or self-hosted NIM microservices handle tasks such as structure prediction, docking and molecule generation. The agent toolkit packages these tasks as skills for agents such as Claude Code and Codex, covering structure prediction, docking, generative chemistry, genomics, protein design and biomarker discovery.1357
- Who needs it?
- Research groups and companies that build or fine-tune biological foundation models, screen or design molecules and proteins computationally, or want AI agents to run computational experiments. Users need Python and PyTorch skills, Docker, and access to recent NVIDIA GPUs for training work (the FP8 recipes need Hopper or newer).
- What does it need?478
- Recent NVIDIA GPUs for training; FP8 recipes need compute capability 9.0 or higher (Hopper and newer), MXFP8 needs compute capability 10.0 or 10.3 (Blackwell) and NVFP4 needs 10.0 or higher
- Docker with GPU support, since each recipe builds and runs as a container
- Python and PyTorch experience; Hugging Face Transformers to load published checkpoints
- An NGC API key for hosted NIM evaluations in the agent toolkit
- A supported coding agent, such as Claude Code or Codex, to use the agent skills
- Your own biological data, or public datasets such as CELLxGENE and UniProt used in the tutorials
- What it is not
- BioNeMo is not a single maintained framework container any more: the BioNeMo Framework container on NGC is archived, and the recipes are self-contained examples rather than a complete training framework. It is not a wet lab or a drug pipeline; model outputs are hypotheses that still need experimental validation. It is also not the same as Parabricks (genomics analysis) or MONAI (a community imaging framework co-founded by NVIDIA and King's College London), which NVIDIA lists as separate platforms.19
Availability and licensing. Much of BioNeMo is published openly, but not every part is open source. BioNeMo Recipes is under Apache-2.0. For the NVIDIA-BioNeMo organization, NVIDIA states that code is generally Apache-2.0, model weights use the NVIDIA Open Model License and data is CC BY 4.0, with exceptions per component. The Agent Toolkit's source code is Apache-2.0, while its skills and documentation are CC-BY-4.0. The BioNeMo Framework container on NGC is archived and no longer maintained (latest tag 2.7.1), and the Framework documentation home says a documentation revamp is in progress. NIM microservices have their own terms.478
The problem it solves
Biological foundation models for proteins, DNA, RNA and cells are large, and training or adapting them on standard code paths is slow and expensive. Many published models also ship as research code that is hard to scale.
At the same time, drug discovery teams want to chain tools such as structure prediction, docking and molecule generation into repeatable workflows, and increasingly want AI agents to run those steps.
BioNeMo responds with GPU-tuned model code and training recipes, served models behind standard APIs, and agent skills that wrap these tools.
How it works
- Pick a model. Model classes compatible with Hugging Face are published with converted checkpoints under the nvidia organization on Hugging Face, for example AMPLIFY and ESM-2.
- Train or fine-tune with a recipe. Each recipe is a self-contained Docker container showing one way to train, using native PyTorch, Hugging Face Accelerate or PyTorch Lightning with TransformerEngine layers, FSDP or Megatron-FSDP, FP8 and context parallelism.
- Serve. NIM microservices expose models such as Boltz-2, DiffDock, Evo 2, GenMol, OpenFold and RFdiffusion behind an API.
- Orchestrate with agents. The BioNeMo Agent Toolkit installs skills into agents such as Claude Code or Codex, so an agent can prepare inputs, call models and run workflows such as a drug discovery pipeline.
- Extend. Libraries such as nvMolKit and cuEquivariance and research models such as La-Proteina, ReaSyn and RNAPro are listed in the NVIDIA-BioNeMo GitHub index; some of them live in their own repositories or organizations.710
Diagram as a list
Applications & solutions
- Biological data (sequences, structures, single-cell data)Input to training and inferenceConnects to BioNeMo Recipes (TransformerEngine, FSDP, FP8), NIM microservices (Boltz-2, OpenFold, Evo 2, GenMol)
- BioNeMo Agent Toolkit skillsLet agents run structure, docking and design workflows
Models & frameworks
- Model checkpoints on Hugging Face (AMPLIFY, ESM-2, Geneformer)Starting points for training and fine-tuningConnects to BioNeMo Recipes (TransformerEngine, FSDP, FP8)
- BioNeMo Recipes (TransformerEngine, FSDP, FP8)Train or adapt biological foundation modelsConnects to Model checkpoints on Hugging Face (AMPLIFY, ESM-2, Geneformer)
- Research models and libraries (La-Proteina, nvMolKit)Additional methods and GPU librariesConnects to BioNeMo Agent Toolkit skills
Inference & runtime software
- NIM microservices (Boltz-2, OpenFold, Evo 2, GenMol)Serve models behind APIsConnects to BioNeMo Agent Toolkit skills
Accelerated computing
- NVIDIA Hopper and Blackwell GPUsRun training and inferenceConnects to BioNeMo Recipes (TransformerEngine, FSDP, FP8), NIM microservices (Boltz-2, OpenFold, Evo 2, GenMol)
Capabilities
Training recipes for biological foundation models4
Self-contained recipes for ESM-2, Geneformer, CodonFM, Llama 3 on genomic data and vision transformers, using TransformerEngine with FSDP or Megatron-FSDP.
Why it matters: Shows a tested, efficient way to train or fine-tune large biology models on NVIDIA GPUs.
Limits: Recipes are examples, not a framework; many features in the support matrix are still under development or unsupported for some models.
Optimized model checkpoints4810
Model classes with TransformerEngine layers for AMPLIFY, ESM-2 and Geneformer, with checkpoints on Hugging Face under the nvidia organization.
Why it matters: Lets teams start from known checkpoints and load them with standard Hugging Face code.
Limits: Precision support differs per model and some cells in the Recipes support matrix are still under development; the older Framework documentation also marks a Llama-3.1 model as in progress.
Low-precision and long-sequence training10
Support for BF16, FP8, MXFP8 and NVFP4 precision, sequence packing that skips padding, and context parallelism for long sequences.
Why it matters: Reduces training cost for long DNA or protein sequences.
Limits: FP8 needs Hopper or newer and MXFP8 and NVFP4 need Blackwell; support differs per recipe.
Agent skills for life science workflows7
The BioNeMo Agent Toolkit packages skills for protein binder design, structure prediction, docking, generative chemistry, ADMET prediction and genomics, installable into Claude Code or Codex.
Why it matters: Lets researchers ask an agent to run multi-step computational experiments with NVIDIA tools.
Limits: In the toolkit's hosted skill evaluations, OpenFold3 and the MSA-Search to OpenFold3 workflow are marked pending validation; agent results need expert review.
Served models through NIM17
NIM microservices for models such as Boltz-2, DiffDock, Evo 2, GenMol, MolMIM, OpenFold2 and OpenFold3, ProteinMPNN and RFdiffusion, which the agent skills call.
Why it matters: Gives standard APIs for common drug discovery steps without hosting research code.
Limits: Hosted evaluations need an NGC API key; check each model's license and terms before production use.
Research models and GPU libraries6
Open research releases in the NVIDIA-BioNeMo organization, such as La-Proteina, Proteina-Complexa, GenMol, RNAPro, KERMT and CodonFM, plus libraries such as nvMolKit and BioNeMo-Inference-Runtime.
Why it matters: Gives access to recent NVIDIA research methods for protein, RNA and molecule design.
Limits: NVIDIA states that, unless otherwise noted, code in these repositories is provided as-is and not actively maintained; licenses differ per component.
Practical use cases
A biotech team wants a protein language model adapted to its own sequence data.
- Approach
- Start from the ESM-2 checkpoint and fine-tune it with the ESM-2 recipe, including the PEFT variant, inside the recipe's container.
- Role of NVIDIA BioNeMo
- Provides the optimized model code and training recipe.
- Data, infrastructure and skills
- Curated sequence data, Hopper or newer GPUs, and ML engineers familiar with PyTorch.
- Type of benefit
- Faster model adaptation
- Caveats
- Fine-tuned models need evaluation on held-out data and on the actual biological question.
- First step
- Build the esm2_native_te recipe container and run its training script on a small dataset.
A drug discovery group wants an agent to run structure prediction, docking and molecule generation as one workflow.
- Approach
- Install the BioNeMo Agent Toolkit skills into Claude Code or Codex and run the drug discovery pipeline skill against NIM-hosted models.
- Role of NVIDIA BioNeMo
- Supplies skills and served models the agent calls.
- Data, infrastructure and skills
- Access to the hosted NIM endpoints the skills call (the toolkit's hosted evaluations use an NGC API key), chemistry and biology experts to review outputs, and data governance rules for what the agent may access.
- Type of benefit
- Shorter cycle between hypothesis and computational result
- Caveats
- Agent and model outputs are hypotheses; some workflows are still pending validation.
- First step
- Install one skill with npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit and run it on a known target.
A research lab needs to train a genomic or single-cell model on long sequences.
- Approach
- Use the Geneformer or Llama 3 on OpenGenome2 recipes with FP8 and context parallelism on a multi-GPU cluster.
- Role of NVIDIA BioNeMo
- Provides tested scaling settings for long-sequence training.
- Data, infrastructure and skills
- A GPU cluster with Hopper or Blackwell GPUs and large training datasets.
- Type of benefit
- Lower training cost
- Caveats
- Precision and parallelism settings must be checked for accuracy loss on your data.
- First step
- Run the Geneformer recipe on a small CELLxGENE subset first.
Works with
Optional integration
- NVIDIA NIMBioNeMo models are served as NIM microservices, and agent skills call NIM-hosted models.
Complementary tools
- NVIDIA NeMoNVIDIA says the Agent Toolkit is built on NIM, Parabricks, NeMo and Nemotron technologies.
- NVIDIA NemotronNemotron is listed among the core pieces of the BioNeMo Agent Toolkit.
- NVIDIA OpenShellOpenShell is listed among the core pieces of the BioNeMo Agent Toolkit.
- NVIDIA BlueprintsNVIDIA publishes a BioNeMo Blueprint for generative protein binder design.
Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Choose your entry point
Decide whether you need training (BioNeMo Recipes), served models (NIM microservices) or agent workflows (Agent Toolkit).
Check: One entry point and one model are chosen.
Load a checkpoint
Load a published model with Hugging Face, for example AutoModel.from_pretrained with nvidia/AMPLIFY_120M.
Check: The model loads and returns embeddings for a test sequence.
Run a recipe
Clone bionemo-recipes, enter recipes/esm2_native_te, build the Docker image and run train.py with --gpus all.
Check: Training starts and logs throughput on your GPU.
Try an agent skill
Install a skill with npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit, targeting Claude Code or Codex.
Check: The agent lists the skill and completes a sample task.
Check licenses
Review the license of each repository, model weight and dataset you use; code, weights and data often differ.
Check: License terms are recorded for every component in your workflow.
Official resources
Could this technology help you?
Describe your project to the Solution Architect. It starts with NVIDIA BioNeMo as context but recommends independently, including when you do not need it.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- NVIDIA Healthcare and Life Sciences (opens in a new tab)
- BioNeMo Framework container (NGC catalog) (opens in a new tab)
- NVIDIA Announces BioNeMo Agent Toolkit (press release, 23 June 2026) (opens in a new tab)
- NVIDIA-BioNeMo/bionemo-recipes repository (opens in a new tab)
- Healthcare and Life Sciences developer resources (opens in a new tab)
- NVIDIA-BioNeMo GitHub organization (opens in a new tab)
- BioNeMo Agent Toolkit repository (opens in a new tab)
- NVIDIA BioNeMo Framework documentation (opens in a new tab)
- London AI Centre and NVIDIA launch MONAI, a new AI framework for healthcare (opens in a new tab)
- BioNeMo Recipes documentation (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.