Protein structure prediction models like AlphaFold's successors have gotten dramatically more accurate over the past few years, but running them at scale is still a GPU engineering problem most biology teams aren't equipped to solve on their own. NVIDIA's answer, announced by NVIDIA Healthcare on September 10, 2026, is BioNeMo Inference Runtime — now in public beta — an open, PyTorch-native library built specifically to make biomolecular inference faster on NVIDIA GPUs.
This is a different layer of the stack than the BioNeMo Agent Toolkit NVIDIA shipped into Claude Science just over a week earlier: that toolkit lets an agent orchestrate a folding workflow from a prompt, while Inference Runtime is what makes the folding models themselves run faster underneath, agent or no agent.
TL;DR
| Question | Answer |
|---|---|
| What is it? | BioNeMo Inference Runtime — an open, PyTorch-native library for faster biomolecular inference on NVIDIA GPUs |
| Status | Public beta, announced September 10, 2026 |
| Which models does it speed up? | Boltz-2, OpenFold2, Protenix v2 |
| How does it speed things up? | Specialized inference kernels and, where supported, CUDA Graphs |
| How does it scale? | Ray-powered GPU replicas for multi-GPU throughput |
| Is it open source? | Yes — PyTorch-native, not a closed runtime |
| How is it different from BioNeMo Agent Toolkit? | Agent Toolkit orchestrates workflows via prompts in Claude Science; Inference Runtime optimizes the raw GPU execution of the models themselves |
What the runtime actually does
BioNeMo Inference Runtime targets the specific pain point of running biomolecular structure-prediction models efficiently: these models involve heavy, irregular computation (attention over large sequence and structure representations) that generic inference stacks don't always execute optimally. NVIDIA's fix is two-pronged:
- Specialized kernels written for the exact operations these models rely on, replacing generic PyTorch ops with GPU-tuned implementations
- CUDA Graphs, where supported, which capture a sequence of GPU operations once and replay them with far less CPU launch overhead — a meaningful win for models that run the same computation pattern repeatedly across many inference calls
Once a single GPU is running as fast as the kernels allow, the runtime hands off scaling to Ray, letting a workload spread across multiple GPU replicas to increase total throughput rather than queuing everything behind one device.
Being PyTorch-native matters for adoption: a team already running Boltz-2 or OpenFold2 in a standard PyTorch environment can plug into the runtime without rewriting their pipeline into a proprietary inference framework — the same reason structured, well-defined tool interfaces matter for agent integrations, familiar interfaces lower the switching cost.
The three models it optimizes
| Model | What it does |
|---|---|
| Boltz-2 | Open-source biomolecular structure and affinity prediction model |
| OpenFold2 | Open-source reimplementation and successor in the AlphaFold-style protein structure prediction lineage |
| Protenix v2 | Structure prediction model in the same class, focused on complex biomolecular assemblies |
These aren't NVIDIA's own models — they're existing open models from the broader structural biology community, which is the point: rather than building a competing model, NVIDIA is building the inference layer that makes the community's existing best models run faster on its hardware.
Why NVIDIA is doing this now
The announcement lists a long roster of collaborators — Apheris AI, Aureka Bio, BioMap, the Boltz team, Dataiku, University of Washington's protein design lab, Latent Labs, MoleculeMind, Nebius, OpenMSF, OpenFold, Proxima Bio, SandboxAQ, Terray Therapeutics, and Xaira Therapeutics — spanning cloud infrastructure, drug discovery startups, and academic labs. That breadth signals NVIDIA positioning itself as the shared inference layer underneath an entire ecosystem of biomolecular AI companies, rather than one product tied to a single partner.
It also lands roughly a week and a half after NVIDIA's BioNeMo Agent Toolkit integration with Claude Science, which raised task correctness on protein-folding workflows from 60% to 100% by giving an agent orchestration skills for MSA generation and OpenFold3/Boltz-2 structure prediction. Read together, the two announcements describe a full stack forming in real time: an agent layer that knows how to sequence a scientific workflow, sitting on top of an inference layer that's now been specifically tuned to run that workflow's models fast.
What people are asking
Do I need Claude Science or an agent to use this? No — BioNeMo Inference Runtime is a standalone inference library. It speeds up the underlying models regardless of whether they're being called by a human script, a pipeline, or an agent like the one in Claude Science.
Does this replace the models themselves? No. Boltz-2, OpenFold2, and Protenix v2 remain separately developed, open-source models; the runtime is purely an execution layer that makes running them on NVIDIA GPUs faster, not a new model architecture.
Is there a benchmark showing the speedup? NVIDIA's announcement doesn't publish specific throughput or latency numbers in the initial post — it describes the mechanism (specialized kernels, CUDA Graphs, Ray scaling) rather than headline benchmark figures. Treat concrete speedup claims as unverified until NVIDIA or an independent team publishes numbers.
What does "public beta" mean here? It means the library is available for external teams to try and give feedback on, but NVIDIA hasn't declared it a finished, stable 1.0 release — expect API changes before general availability.
The bottom line
BioNeMo Inference Runtime is NVIDIA extending its usual playbook — build the open-source ecosystem's favorite models a faster place to run — into structural biology specifically. Pairing it with the recent BioNeMo Agent Toolkit / Claude Science integration suggests NVIDIA is building both ends of the same pipe at once: the agent layer that knows what to run, and the inference layer that runs it fast. For teams already using Boltz-2, OpenFold2, or Protenix v2 in production, it's worth testing against an existing pipeline now that it's in public beta.
Details reflect NVIDIA Healthcare's September 10, 2026 announcement; check the official announcement for benchmark data and API changes as the beta progresses.
Related reading
- NVIDIA BioNeMo Agent Toolkit comes to Claude Science: protein prediction by prompt
- Google Research ToolGrad: answer-first tool-use dataset generation
- Tool definition and schema design: the context engineering layer most teams get wrong
- Can AI cure cancer?
- NVIDIA Cosmos 3: open physical AI world model guide
- What are AI agents? A complete guide
- Official: NVIDIA Healthcare — BioNeMo Inference Runtime announcement