Chai-1

Chai-1 is a multimodal foundation model for biomolecular structure prediction developed by Chai Discovery, released in September 2024 with technical report posted to bioRxiv in October 2024. The model produces unified all-atom predictions for complexes of proteins, small molecules, DNA, RNA, and post-translational modifications, and can be conditioned on experimental restraints to improve accuracy on tasks where partial structural data is already available. As of its release, Chai-1 was the first AlphaFold-class structure prediction system released by a venture-backed AI startup as both a free hosted web tool for commercial drug discovery use and an open weights download for non-commercial research.

At a glance

• Lab: Chai Discovery
• Released: September 9, 2024 (initial release); technical preprint posted to bioRxiv on October 10, 2024.
• Modality: Biomolecular structure prediction. Inputs include protein sequences, nucleic acid sequences, small molecule ligands (SMILES), and optional experimental restraints. Outputs are predicted three dimensional all-atom structures.
• Open weights: Yes, partial. Model weights and inference code are released as a Python package on GitHub and Hugging Face for non-commercial research use. Commercial use through the hosted web tool is permitted, including for drug discovery.
• Context window: Not applicable in the language model sense. The model handles biomolecular complexes spanning proteins, ligands, and nucleic acids in a single inference pass.
• Pricing: Free for the hosted web tool at chaidiscovery.com, including for commercial drug discovery. The open weights distribution is free for non-commercial research; commercial deployment of the local weights requires a separate license arrangement.
• Distribution channels: chaidiscovery/chai-lab on GitHub, chaidiscovery/chai-1 on Hugging Face, and the hosted web tool at chaidiscovery.com.

Origins

Chai Discovery was founded in March 2024 in San Francisco by Joshua Meier, Jack Dent, Matthew McPartlon, and Jacques Boitreaud. Meier had been a research and engineering staff member at OpenAI from 2018 and went on to lead machine learning at the protein design firm Absci before co-founding Chai. Dent had been a software engineer at Stripe and is a Harvard classmate of Meier. McPartlon and Boitreaud both came from structural biology and machine learning research backgrounds. The company's seed round of $30 million closed in September 2024 with participation from Thrive Capital, OpenAI, and Dimension, and the Chai-1 release was the company's first major public capability disclosure.

The release sequence positioned Chai-1 as a research credibility anchor for a company built on a longer commercial drug discovery thesis. The model was published with a free public web tool, an open weights download for non-commercial use, and a technical report on bioRxiv. Industry coverage at the September release characterized the move as deliberate community building, comparable to the early Hugging Face strategy of seeding open distribution before monetizing the surrounding platform.

The architectural lineage traces to AlphaFold 2 and AlphaFold 3, the Google DeepMind and Isomorphic Labs structure prediction systems that defined the modern category, and to the ESM protein language model line from Meta AI / FAIR and EvolutionaryScale. Chai-1 differentiated by moving beyond protein only structure prediction to unified all-atom prediction across proteins, small molecules, nucleic acids, and modifications, and by supporting experimental restraint conditioning that improves accuracy when partial structural data is available.

The technical report described training on the Protein Data Bank and additional structural datasets, with the model architecture broadly aligned with the AlphaFold 3 publicly described approach but trained independently and released by a venture-backed startup rather than by a research institution.

Capabilities

Chai-1 handles unified all-atom structure prediction for biomolecular complexes spanning proteins, small molecules, DNA, RNA, and post-translational modifications. The model produces three dimensional structures from sequence and ligand inputs, and the unified approach is the principal architectural distinction from earlier protein only systems.

Three capability features distinguish Chai-1 from peer structure prediction systems.

The first is multimodal coverage. Chai-1 accepts protein sequences, nucleic acid sequences, small molecule ligands as SMILES strings, and metal ions or modifications in a single inference call. The output is a unified all-atom structure rather than separate protein and ligand predictions stitched together. The unified approach matters for drug discovery applications, where the structural question is typically about the protein in complex with a candidate small molecule, not the protein alone.

The second is restraint conditioning. The model can be prompted with experimental restraints, including known contacts or partial structures, to refine predictions when partial laboratory data is already available. The restraint conditioning is particularly useful for antibody engineering tasks where small amounts of cross-linking, NMR, or hydrogen-deuterium exchange data can substantially improve prediction accuracy.

The third is single-sequence operation. Chai-1 can run without multiple sequence alignment (MSA) input, which is the standard input mode for AlphaFold 2 and many peer systems. Running single-sequence preserves most of the model's structure prediction performance while eliminating the MSA generation step, which is computationally expensive and requires access to large protein sequence databases. Single-sequence operation makes the model more accessible to researchers who lack the infrastructure for MSA generation pipelines.

Benchmarks and standing

Chai-1 was characterized in the technical report and in industry coverage as state of the art on multiple structure prediction benchmarks at its September 2024 release.

On the CASP15 protein monomer structure prediction set, Chai-1 reported a Cα LDDT (Local Distance Difference Test) score of 0.849, surpassing the ESM3-98B model from EvolutionaryScale at 0.801. CASP is the principal community benchmark for protein structure prediction, run as a biennial blind test in which structures held out of training are predicted by participating systems and compared against subsequently solved experimental structures.

On protein-ligand structure prediction tasks, Chai-1 reported success rates competitive with AlphaFold 3 from Isomorphic Labs on standard evaluation sets, while maintaining the unified all-atom architecture across additional molecule classes that earlier systems did not handle directly.

On antibody-antigen structure prediction, the model's restraint conditioning was characterized as a particularly useful capability for accelerating antibody engineering workflows where small amounts of experimental data are typically available.

The horizontal language model benchmarks (Artificial Analysis Intelligence Index, LMArena, GPQA Diamond, AIME, SWE-bench) do not apply to biomolecular structure prediction. The relevant benchmarks are CASP, the PDB based evaluation sets used in the AlphaFold publications, and antibody design hit rate evaluations of the form later associated with Chai-2.

Industry coverage in 2024 and 2025 has consistently grouped Chai-1 with AlphaFold 3 and ESM3 as the principal frontier biomolecular structure prediction systems, with the open weights and free hosted tool distribution as Chai's distinctive commercial positioning.

Access and pricing

Chai-1 is accessible through three surfaces. The hosted web tool at chaidiscovery.com is free, including for commercial drug discovery applications, and was the principal launch surface in September 2024. The Python package and model weights are available at chaidiscovery/chai-lab on GitHub and at chaidiscovery/chai-1 on Hugging Face under terms permitting non-commercial research use. Commercial deployment of the local weights requires a separate license from Chai Discovery.

There is no per-token or per-prediction billing on the hosted tool as of the release configuration. Chai Discovery's commercial revenue model centers on direct partnerships with pharmaceutical companies (the Eli Lilly collaboration announced January 2026) and on closed source successor models (Chai-2) rather than on hosted Chai-1 inference.

Comparison

Direct competitors and adjacent biomolecular structure prediction systems:

• AlphaFold 3 (Isomorphic Labs, Google DeepMind). The closest research peer. AlphaFold 3 was published in May 2024 with comparable unified all-atom architecture covering proteins, ligands, and nucleic acids. AlphaFold 3 was initially released through the AlphaFold Server with strict use restrictions, while Chai-1 was released with broader commercial use permissions on the hosted tool and open weights for non-commercial research, which positioned Chai as the more accessible option for academic and startup users at the September 2024 release.
• ESM3 (EvolutionaryScale). Multimodal protein language model with structure prediction capability. ESM3 spans sequence, structure, and function reasoning rather than focusing primarily on structure prediction. Chai-1's CASP15 result reportedly led ESM3-98B on protein monomer structure accuracy.
• AlphaFold 2 (Google DeepMind). The structural prior generation system. Still widely used, with broad community tooling and open source predecessors. Chai-1 supersedes AlphaFold 2 on most accuracy benchmarks and on the unified all-atom front.
• OpenFold and the academic structure prediction lineage. OpenFold and adjacent open source structure prediction projects continue to provide alternatives at lower complexity, particularly for users who need a fully open source training pipeline rather than only inference weights.
• Boltz-1 (academic). A subsequent open source biomolecular structure prediction model released by MIT researchers under permissive licensing, often compared to Chai-1 in technical reviews of the late 2024 and 2025 structure prediction landscape.

Chai-1's distinctive position among 2024 vintage biomolecular structure prediction systems: state of the art accuracy on CASP15, unified all-atom prediction across proteins and ligands and nucleic acids, restraint conditioning that improves antibody engineering applications, and a distribution model that combined a free commercial hosted tool with open weights for research.

Outlook

Open questions for Chai-1 over the next 6 to 18 months:

• Continued open access. Chai Discovery has shifted toward closed source for the Chai-2 successor and toward bespoke partner models such as the Eli Lilly collaboration. Whether Chai-1 receives further updates as an open distribution release will signal the company's continued investment in the open research surface that anchored its September 2024 launch.
• Adoption in academic and startup pipelines. Chai-1 is already widely used in research workflows. The longevity of that adoption will depend on whether AlphaFold 3 and successor open releases close the accessibility gap, and whether Boltz-1 and adjacent open source efforts produce competitive successors at lower commercial friction.
• Restraint conditioning benchmarks. The experimental restraint capability is the most distinctive Chai-1 feature for antibody engineering and structural biology. The community development of standardized restraint conditioning evaluations would clarify how much accuracy uplift the feature provides on representative pharmaceutical workflows.
• Comparison data versus Chai-2. The 2025 release of Chai-2 shifted Chai's flagship effort to zero-shot antibody design rather than structure prediction. Public comparisons of structure prediction accuracy between Chai-1 and Chai-2 (where applicable) would help users decide which model serves which workflow.

Sources

• Chai-1: Decoding the molecular interactions of life (bioRxiv). Technical report covering architecture, training data, and benchmark results.
• GitHub: chaidiscovery/chai-lab. Open source Python package and inference code.
• Hugging Face: chaidiscovery/chai-1. Model weights and model card.
• Chai Discovery blog: Introducing Chai-1. Official launch announcement.
• MarkTechPost: Chai-1 Released by Chai Discovery Team. Release coverage.
• LabCritics: Chai-1, the Next Frontier in Molecular Structure Prediction. Independent technical review.
• BioLM: Chai-1 model card. Capability and benchmark summary.