Tasks and inputs
BioEmu-2 is listed for Any-to-any, Data & control. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
BioEmu-2 is Microsoft Research's open ~500M diffusion model that samples ensembles of all-atom biomolecular structures, including complexes and ligands.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
A single predicted structure hides the fact that proteins move. BioEmu-2 samples ensembles instead: many plausible all-atom 3D conformations for a biomolecule, so you can study how it flexes and binds rather than staring at one frozen pose.
The first BioEmu dealt with protein chains. This version handles multi-chain complexes, ligands and nucleic acids as well.
| Field | Value |
|---|---|
| Repository | microsoft/bioemu2 |
| Developer | Microsoft Research |
| Architecture | Diffusion-based all-atom score model |
| Size | About 500M parameters |
| Inputs | Sequence(s), optional MSA, contact conditioning, method conditioning |
| Outputs | Ensemble of all-atom structures in Cartesian coordinates |
| Trained | August–September 2026 |
| Released | October 2026 |
| Licence | MIT |
Contact conditioning lets you steer sampling with distances you already know between particular atoms, from crosslinking or NMR for instance.
The limits are spelled out and worth repeating. It does not generate new sequences. It does not predict binding affinities or absolute free energies. It is not for clinical or diagnostic decisions, nor for designing hazardous biological agents. Every structure is a computational hypothesis until validated.
Quality is expected to fall with shallow or poor MSAs, which covers orphan proteins, engineered constructs and some viral proteins. Training data over-represent well-studied families. Microsoft calls it a research model and does not recommend commercial use without further testing.
The smallest device Microsoft tested is an 80 GB A100. Inference code includes a Python API and a CLI.
Pick three systems with known alternative states, perhaps a kinase with active and inactive structures in the PDB. Sample a few hundred conformations for each and check whether the ensemble covers both states, and in what ratio. Then repeat with a deliberately shallow MSA to see how fast quality degrades for your class of targets.
For design rather than ensemble sampling, NVIDIA and the Baker Lab's RFD4-Proteína is the complement. More Microsoft models are on its provider page.
Compare the model on the work you need to complete.
BioEmu-2 is listed for Any-to-any, Data & control. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
RFD4-Proteína is an open Baker Lab and NVIDIA flow-matching model that designs protein sequence and all-atom structure together, under BSD-3-Clause.