Chai Discovery

The model is the product. Chai licenses foundation models for molecular design to pharmaceutical and life sciences organizations rather than using them to build its own pipeline: the customer brings a target, a chosen epitope and a required antibody format, and the model designs the binder, with custom versions trained on a customer's proprietary data offered as part of at least one announced agreement. There is no wet lab franchise, no instrument line and no therapeutic pipeline underneath. Remove the models and there is nothing to license.

The product designs candidates that the customer then makes and tests: the company's own account has designs going straight to characterization within two weeks instead of high throughput screening, and the reported results are experimental outcomes on tested designs rather than model scores. Access is controlled rather than open, granted on application under a stated deployment policy, and an acceptable use policy binds what may be done with the models, their outputs and derivatives, with suspension for violation and a prohibition on attempts to circumvent safety filters.

No confidence measure is exposed with a value at which a design should not be pursued, and no guidance says where the models are unreliable beyond the stated difficulty with flexible loop regions. The product page's claim that the design "works reliably" is a quality statement rather than a promise about every output.

The models are identified and versioned, without the training data characterized. Three generations are named with a public release history and stated task scope: Chai-1 (2024) for structure prediction, released with weights under a community license; Chai-2 (2025) for de novo antibody and miniprotein design; and Chai-3 (2026), which the company describes as roughly doubling its predecessor's success rate. Technical results are published in the company's own research updates and technical reports describing method and test design, so which model produced a result, and what changed between generations, can be traced.

No characterization of what the models were trained on is published, and there is no model card for any of them. The technical report behind the latest results may say more about training.

The chain is short: the models are Chai's own, named and versioned with a public release history (Chai-1 for structure prediction, Chai-2 for de novo design, Chai-3), and Chai-1 was released with weights under a community license, so for that model a customer holds the artifact itself. Only Chai's own models sit between a customer's target and an answer, and at least one announced agreement includes a custom model trained on the customer's proprietary data.

Its trust center names the subprocessors around them: Amazon Web Services as cloud provider, GitHub Enterprise for code and model files, Google Workspace for single sign on and Vanta for compliance monitoring.

Vendor run studies with method and denominators, named production customers, and no candidate in the clinic. The November 2025 research update sets out the method and the counts: 88 designed IgGs across 28 target antigens tested on four developability criteria using thresholds from named prior analyses of therapeutic antibodies, 86 percent with zero or one flagged issue, 24 of 28 antigens with at least one clean design, cryo-EM confirmation of the intended epitope on five complexes, binders obtained for all six GPCRs tested from 10 to 73 designs each, and two hits against a KRAS G12V peptide-MHC complex that do not bind the G12D or wild type peptides. The product page states double digit success rates for antibodies and above 50 percent for miniproteins, and characterization within two weeks without high throughput screening.

Adoption is named: license and collaboration agreements with Eli Lilly, Pfizer, Novartis and argenx are announced on the company's own pages. The work is the company's own and has not been independently replicated, no peer reviewed paper is named, and no molecule designed with these models is reported in a clinical study.

Chai's privacy policy (last updated 30 June 2025) covers personal information collected through the website and the company's online offerings, with retention stated as a principle rather than a schedule, and its acceptable use policy prohibits using the models to violate anyone's privacy rights. The material that matters most is the customer's scientific data: the targets, structures and sequences submitted to the hosted models, and the proprietary datasets behind a custom model the company trains for a customer.

Whether submissions or customer datasets improve models other customers use is not stated.

A privacy policy covers the service. The privacy policy (last updated 30 June 2025) covers the website and the company's online offerings, describing what personal information is collected, how it is used and how long it is kept, and the site terms bind the organization whose employee accepts them.

Nothing says whether Chai acts as a business associate or whether the models may receive protected health information, which is a live question for a product whose inputs can be a target structure or sequence derived from patient material and whose custom models are trained on a customer's proprietary datasets.

Chai runs a trust center that names SOC 2 Type I and ISO 27001:2022 compliance, lists an ISO 27001 certificate and a SOC 2 Type II report among its compliance documents, monitors its controls continuously through Vanta, and publishes policies on access control, asset management and business continuity. Its customers are large pharmaceutical organizations sending target structures and, in at least one announced agreement, proprietary datasets for custom model training.

The reports and the certificate are released only after a request for access.

Chai licenses preclinical design models to research organizations; nothing in its published material presents them as regulated products or makes a clinical claim, and the company does not claim downstream development capability. A molecule designed with the models carries its own regulatory status, held by the organization that develops it.

The company states an aspiration to generate candidates ready for IND enabling studies in a single run, which describes where it wants the technology to reach rather than a position on what it sells today.

Results are published by the groups the product serves, with counts, alongside a binding use policy. The November 2025 research update reports Chai-2's designs broken out by target class rather than as one headline rate: 88 designed IgGs across 28 target antigens tested against four developability criteria using thresholds from named prior analyses of therapeutic antibodies, with 86 percent of designs showing zero or one flagged issue and 24 of 28 antigens yielding at least one clean design, so the four that did not are on the record; six GPCRs with 10 to 73 designs tested each and at least one binder for every target; and three peptide-MHC complexes with 27 to 50 designs each, yielding two hits on a KRAS G12V complex. Negative controls are published with them: those binders do not bind the G12D or wild type peptides, which reports what the model correctly declines to bind. A known limitation is stated in the company's own materials, that modeling the flexible loop regions of antibodies remains harder than more rigid scaffolds.

Governance of release sits beside that. An acceptable use policy identifies prohibited uses of the models, their outputs and derivatives, is incorporated into the license under which weights are distributed and into the application terms, and permits suspension for violation; access to the design models is granted on application under a stated responsible deployment policy rather than by open download. What is missing: no model card for any model, and the deployment policy's criteria and decision process are not published, so an applicant cannot know what is restricted or who decides.

Success rates are published with their denominators, and no remedy attaches to a design that fails. The November 2025 research update reports what the product achieves and on how many attempts: 88 designed IgGs across 28 antigens assessed on four developability criteria against thresholds from named prior analyses, with 86 percent showing zero or one flagged issue and 24 of 28 antigens yielding a clean design; six GPCRs with 10 to 73 designs tested each and a binder obtained for every one; and three peptide MHC complexes with 27 to 50 designs each, yielding two hits. Cryo-EM on five designed complexes is reported to confirm the intended epitope at sub angstrom accuracy. Those are the company's own experiments rather than an outside evaluation, and each figure comes with its count.

The site terms provide the services as is without warranties of any kind and require the customer to indemnify the company in connection with its use of output, and no remediation or fee consequence attaches when a designed molecule fails.

Research organizations license the design models, so the record systems that matter to Chai's buyers are research systems such as electronic lab notebooks, laboratory information systems and cloud data environments. The design models are reached through access granted on application, the inputs are a structure or a sequence supplied by the customer, and the one openly distributed model is downloaded and run by the customer itself.

No integration with those research systems is offered.

Access routes are stated and residency is not. Three are described: the structure prediction model released with weights under a community license, which a customer runs wherever it chooses; hosted access to the design models, granted on application rather than by open API; and custom models the company trains on a customer's proprietary data under an agreement. That is more than one option and it is the hosted path that carries a customer's targets.

The trust center names Amazon Web Services as the cloud provider, but not the region where a submitted structure or sequence rests.

Access is described in some detail; price is not. Commercial use of the design models runs through an access form that asks for organization type and size and how the applicant plans to use the model, followed by a call; academic and non profit use is described as limited non commercial access through the same form, and the earlier structure prediction model was released with weights under a community license, which is a stated position rather than a price. Announced agreements with Eli Lilly, Pfizer, Novartis and argenx are published without financial terms, and at least one is described as including early access to a next generation model and a custom model trained on the customer's data.

What is sold is clear, a model license rather than a service or a molecule; what it costs, and what drives the cost, is not.

The scope is narrow and named precisely, with results reported in each part of it. The product is preclinical molecular design for pharmaceutical and life sciences research: antibodies in full monoclonal and VHH formats and miniproteins, designed against a target and epitope the customer specifies, with stated support for choosing an epitope, designing for agonism or antagonism against a particular antigen state, handling post translational modifications such as glycans on the epitope, and controlling species cross reactivity. The company does not claim clinical or downstream development capability.

Testing runs across the classes it names: 28 target antigens in the developability work, six G protein coupled receptors, a class the company notes is hard for screening methods, and three peptide MHC complexes, each with the number of designs tested reported. Coverage claims outside protein binders are not made.