Chai-3

Chai-3 is a generative foundation model for biomolecular design developed by Chai Discovery, and the third generation in the company's lineage of biomolecular AI models after Chai-1 and Chai-2. Where Chai-1 focused on predicting the three-dimensional structures of proteins, ligands, nucleic acids, and glycans, the later generations turn that structural understanding toward generative design — producing new molecules, most prominently therapeutic antibodies, rather than only modeling existing ones.

The model was revealed publicly through a licensing agreement with Pfizer, under which Pfizer became one of the first pharmaceutical partners to gain access to Chai-3. It is positioned as a step-change over its immediate predecessor, Chai-2, which was the first zero-shot antibody design platform to reach double-digit experimental hit rates and produce molecules with drug-like properties — a roughly hundredfold improvement over prior computational approaches. Chai-3 is reported to double that success rate while extending the range of targets it can address.

Chai-3 sits within a broader shift in the field from structure prediction toward generative therapeutic design, where the goal is not to model a known complex but to propose novel binders that succeed when synthesized and tested in the laboratory. It targets some of the hardest problems in that space: multispecific molecules and traditionally undruggable targets.

#Key Features

• Zero-shot antibody design: Generates candidate antibodies against therapeutic targets without target-specific training, continuing the zero-shot design paradigm established by Chai-2.
• Doubled success rate: Reported to double the experimental success rate of Chai-2, producing a larger fraction of designs that meet therapeutic standards when validated in the wet lab.
• Multispecific design: Designs molecules with multiple binding specificities, a capability central to next-generation antibody therapeutics.
• Hard-to-drug targets: Advances design against targets that have historically resisted both small-molecule and biologic approaches.
• Improved generalization: Extends reliable performance across a wider set of targets and design scenarios than earlier generations.

#Technical Details

Chai-3 is a multimodal generative model in the Chai lineage, building on the structural modeling foundation established by Chai-1, an open-weights diffusion model that jointly represents proteins, small molecules, DNA, RNA, and glycans. Chai-2 (bioRxiv 2025.07.05.663018) demonstrated the generative direction of this lineage, achieving double-digit experimental hit rates in zero-shot antibody design across dozens of targets. Chai-3 is reported to roughly double Chai-2's success rate and to broaden its target coverage, particularly for multispecifics and difficult targets.

No technical report or preprint describing Chai-3 has been released, and the model's exact architecture, parameter count, training data, and formal benchmark results have not been published. All public technical characterizations trace to the announcement of the Pfizer licensing agreement rather than to a peer-reviewed or preprint source.

#Applications

Chai-3 is aimed at therapeutic discovery, especially the design of antibodies and multispecific biologics against targets of pharmaceutical interest. Its primary demonstrated deployment is within Pfizer's drug discovery operations, where it is used alongside a custom model tuned on Pfizer's proprietary data and workflows. The emphasis on hard-to-drug targets and multispecifics positions it for programs where conventional discovery pipelines struggle to produce viable binders.

#Impact

Chai-3 reflects the accelerating adoption of frontier generative models by major pharmaceutical companies, moving biomolecular AI from research demonstrations into embedded use across real discovery workflows. Its reported gains over Chai-2 extend a rapid generational cadence within Chai Discovery's antibody design program. Important caveats accompany this: unlike Chai-1, which is released under an open license with public weights and code, Chai-3 has no public weights, code, or hosted API, and access appears limited to commercial partners. No model card, data card, or technical report currently exists, so its capabilities are known only through the developer's own announcement rather than independent evaluation.