What is Protenix-v2?
Protenix-v2 is ByteDance Seed’s enhanced-capacity model for biomolecular structure prediction and design. It is designed for scientific workflows in which researchers need to estimate how biological molecules may be arranged in three dimensions. Its documented scope includes proteins, nucleic acids, ligands, and larger molecular complexes.
The model is an evolution of the Protenix structure-prediction system. ByteDance Seed released Protenix-v2 on April 8, 2026, and describes it as a higher-capacity version intended to improve difficult structure-prediction and biomolecular-design tasks. The model is not a general-purpose language model: it does not produce conversational answers, standard programming assistance, images, audio, or video.
In practical terms, Protenix-v2 is most relevant to computational structural biology, antibody research, and early-stage drug-discovery workflows. Its predictions can help researchers formulate hypotheses about molecular structures, but they do not replace laboratory validation or establish experimental or clinical success.
Where Protenix-v2 fits and how it works
Protenix-v2 belongs to ByteDance Seed’s Protenix family and is listed as the current research model in the supplied documentation. The public Protenix project provides software and inference tooling, while access to the Protenix-v2 weights is restricted. This creates an important distinction between the availability of the surrounding research software and the availability of the model itself.
The documented model has approximately 464.44 million parameters. Parameters are the learned numerical values that store patterns acquired during training; the parameter count gives a rough indication of model scale but does not, by itself, guarantee prediction quality. Protenix-v2 uses a representation dimension of c_z=256, ten recycling cycles, and 200 diffusion steps in the documented configuration.
Recycling means that the system repeatedly revisits and refines an intermediate structural prediction. Diffusion steps refer to the gradual sampling process used to produce a molecular structure from a noisy representation. More refinement steps can support difficult predictions, although they also affect computational cost and runtime.
The official documentation lists support for multiple-sequence alignments, RNA multiple-sequence alignments, and templates. A multiple-sequence alignment, or MSA, compares related biological sequences to reveal evolutionary patterns that can help constrain a structure prediction. Templates provide known structural information that may also guide the model when suitable examples are available.
Main capabilities
- Protein structure prediction: Protenix-v2 can be used in workflows that estimate the structures of proteins and protein-containing complexes.
- Nucleic-acid modeling: The documented inputs include RNA MSA support, making the model relevant to RNA and other nucleic-acid research workflows.
- Ligand-aware complex modeling: The model is intended for complexes that include ligands, proteins, nucleic acids, or combinations of these components.
- Antibody-antigen modeling: Reported evaluations emphasize difficult antibody-antigen prediction tasks, an area where small structural differences can be scientifically important.
- Computational biomolecular design: The technical report describes applications involving VHH-Fc campaigns, difficult GPCR targets, and mini-binder design for SARS-CoV-2 spike RBD variants.
Protenix-v2 also adds training-free guidance during diffusion sampling. This allows geometric and physical constraints to influence sampling without retraining the underlying model. The documented guidance mechanisms include steric, torsion, bond, and pairwise-distance potentials. In accessible terms, these constraints can discourage implausible clashes or geometries and help steer sampling toward structures that satisfy selected physical relationships.
What performance has been reported?
ByteDance Seed reports that Protenix-v2 improves antibody-antigen structure prediction compared with Protenix-v1. Across three reported collections, the provider describes absolute success-rate gains of up to 13 percentage points at the stated DockQ threshold. DockQ is a structural-comparison measure used to assess how closely a predicted protein interaction resembles a reference structure.
The project documentation also reports that five Protenix-v2 sampling seeds can exceed the performance achieved by Protenix-v1 with 1,000 seeds for the evaluated tasks. A sampling seed is a different starting condition for stochastic generation. If the reported result generalizes to a researcher’s workload, it could reduce the number of repeated samples needed to obtain a strong candidate prediction.
These figures are provider-reported research results, not an independent guarantee of performance on every target. The accompanying technical report also describes design experiments involving novelty-controlled VHH-Fc campaigns, difficult GPCR targets, and SARS-CoV-2 spike RBD mini-binders. Such results show the research areas tested by the authors, but they should not be treated as proof that a proposed molecule will work in a laboratory, animal study, or clinical setting.
Technical specifications and supported inputs
| Specification | Documented information |
|---|---|
| Provider | ByteDance Seed |
| Model family | Protenix |
| Model type | Biomolecular structure prediction and design |
| Approximate parameter count | 464.44 million |
| Representation dimension | c_z=256 |
| Recycling cycles | 10 |
| Diffusion steps | 200 |
| Documented input features | MSA, RNA MSA, and templates |
| Training-data cutoff | September 30, 2021 |
| Release date | April 8, 2026 |
The September 30, 2021 date is a scientific training-data cutoff identified in the model documentation. It should not be interpreted as a conventional language-model knowledge cutoff, because Protenix-v2 is not a conversational model and does not answer questions by retrieving text from a general knowledge corpus.
No official context-window limit, maximum output-token limit, token-based pricing, hosted API specification, streaming interface, or batch API specification was identified in the supplied research. These omissions matter for planning: users should not assume that Protenix-v2 can be called like a commercial language model or that it has a standard text API with published token limits.
Availability, licensing, and limitations
The Protenix repository and associated inference tooling are publicly available. However, ByteDance Seed identifies the Protenix-v2 weights as proprietary and confidential. They are not released under an open-source license and may not be reproduced, distributed, sublicensed, disclosed, or transferred without permission from the rights holder.
This licensing arrangement is one of the model’s most significant practical limitations. A research team may be able to inspect and use the public software project, but that does not mean it can freely download, modify, redistribute, or host the Protenix-v2 weights. Organizations should review the applicable permissions before building a production workflow or sharing derived infrastructure.
Protenix-v2 also requires specialist preparation. Depending on the task, users may need to prepare biological sequences, generate MSAs, provide templates, configure diffusion sampling, and supply appropriate computational infrastructure. The model is therefore less suitable for casual users or teams seeking an immediately accessible hosted application.
There are no documented capabilities for conversational text generation, general-purpose coding, standard tool or function calling, web search, structured text output, image generation, audio generation, or video generation. The research records no official pricing, so cost cannot be compared directly with per-token commercial AI services. In practice, the total cost will depend on access arrangements, hardware, preprocessing, sampling configuration, and the computational demands of the research workflow.
When to choose Protenix-v2
Protenix-v2 is a strong candidate when the central problem is biomolecular structure prediction rather than language generation. It is particularly relevant when a project involves one or more of the following:
- Predicting protein, RNA, nucleic-acid, ligand, or multi-component complex structures.
- Studying antibody-antigen interactions or other difficult molecular interfaces.
- Exploring computational binder or biomolecular design candidates before laboratory testing.
- Using diffusion-based structure prediction with geometric or physical constraints.
- Evaluating a research workflow that benefits from multiple-sequence alignments, RNA MSAs, and templates.
The reported sampling-efficiency result may also appeal to teams that need to explore several structural candidates. According to ByteDance Seed’s documentation, five Protenix-v2 seeds exceeded the evaluated performance of Protenix-v1 with 1,000 seeds. This is a provider-reported comparison, so teams should validate the trade-off on their own target classes before assuming equivalent efficiency.
When another option may be more appropriate
A different option may be preferable if the project requires an openly licensed set of model weights, a managed hosted endpoint, predictable token-based pricing, or a conventional API with published context and output limits. Protenix-v2 is also not the right choice for chat, document processing, coding assistance, image creation, speech, or video generation.
Teams with limited computational biology experience may need a more integrated workflow that handles sequence preparation, MSA generation, templates, and infrastructure. Conversely, specialist structural-biology groups may value the public software project and the model’s research focus even though the weights are proprietary. The choice depends less on generic model size than on whether the team can satisfy the biological-input, licensing, and compute requirements.
Overall assessment
Protenix-v2 is a focused research model for biomolecular structure prediction and design. Its defining characteristics are its approximately 464.44-million-parameter scale, support for protein and nucleic-acid-related inputs, diffusion-based sampling, training-free physical guidance, and reported gains on antibody-antigen tasks. It is not a general AI assistant and should not be evaluated using conversational quality, coding benchmarks, or token pricing.
For researchers working on structural biology or computational biomolecular design, Protenix-v2 offers a specialized capability set and a public surrounding software project. The main qualifications are equally clear: the weights are proprietary, no hosted pricing or standard API limits are documented, and the reported scientific results require independent validation for each intended use case.