Chai-2

Chai-2 is a multimodal generative model for zero-shot antibody and binder design developed by Chai Discovery, released in 2025 as the company's transition from open source structure prediction (Chai-1) toward closed source generative drug design products. The model produces fully de novo antibody complementarity-determining regions (CDRs) from an input epitope and target structure, without reliance on pre-existing antibody templates or training-time scaffolds. Chai Discovery reported experimental hit rates of 16 to 20 percent across 52 previously unaddressed antigen targets in a 24-well plate format, a result the company has characterized as more than 100 times the success rate of prior computational methods and the first published zero-shot antibody design platform to clear the double digit hit rate threshold.

At a glance

• Lab: Chai Discovery
• Released: Public preprint and announcement July 5, 2025; updated bioRxiv preprint posted November 2025.
• Modality: Generative biomolecular design. Inputs include target antigen structure and a specified epitope. Outputs are fully de novo antibody CDR designs and binder candidates.
• Open weights: No. Chai-2 is closed source. Access is restricted to commercial partner deployments and to the company's own internal pipeline.
• Context window: Not applicable in the language model sense. The model handles antibody design tasks specified by target structure and epitope.
• Pricing: Not publicly disclosed. Chai-2 is not offered as a hosted general access tool. Access is through partnership arrangements with pharmaceutical companies (the January 2026 Eli Lilly collaboration is the most prominent disclosed example).
• Distribution channels: Direct partnerships and Chai Discovery's internal molecule design pipelines. The model is not available through a public API or open weights distribution.

Origins

Chai Discovery was founded in San Francisco in March 2024 by Joshua Meier (former OpenAI staff and former Absci machine learning lead), Jack Dent (former Stripe engineer), Matthew McPartlon, and Jacques Boitreaud. The company's first public model, Chai-1, was released in September 2024 as an open source structure prediction system positioned alongside AlphaFold 3 from Isomorphic Labs and ESM3 from EvolutionaryScale. Chai-2 was the company's second model release and the first generation of a closed source generative design line.

The Chai-2 release on July 5, 2025, took the form of a bioRxiv preprint titled "Zero-shot antibody design in a 24-well plate" and an accompanying announcement on the Chai Discovery website. The release framed the work as a category transition: Chai-1 had been positioned as a structural prediction tool, but Chai-2 was positioned as a generative design platform whose outputs are validated by laboratory experiments rather than by computational consistency metrics alone. The 24-well plate framing in the paper title refers to the standard low-throughput experimental format, and the company's claim was that Chai-2 produced sufficiently high quality antibody designs that meaningful screening campaigns could be run in a single 24-well plate rather than the high-throughput parallel screening (millions of candidates) that traditional antibody discovery requires.

A subsequent updated preprint posted to bioRxiv in November 2025 ("Drug-like antibody design against challenging targets with atomic precision") extended the Chai-2 capability disclosure to harder target classes, with continued reporting of double digit zero-shot hit rates. The November 2025 release was paired with the company's $130 million Series B at a $1.3 billion post-money valuation, co-led by Oak HC/FT and General Catalyst, and the January 2026 announcement of a multi-target collaboration with Eli Lilly on biologics discovery.

The shift from Chai-1's open weights distribution to Chai-2's closed source posture was a deliberate strategic move. Industry coverage at the time of the Series B characterized the transition as Chai Discovery's pivot from research credibility seeding (Chai-1) to commercial revenue generation through pharmaceutical partnerships, with the closed source posture protecting the model's value to those partnerships.

Capabilities

Chai-2 is built specifically for zero-shot generative design of antibodies and binders. The model takes a target antigen structure and a specified epitope as input and produces fully de novo antibody CDR designs as output, without relying on pre-existing antibody templates or training-time scaffolds for the target.

Three capability features distinguish Chai-2 from peer antibody design systems.

The first is zero-shot operation. The model does not require fine-tuning on the specific target before generating designs. Traditional antibody discovery typically requires either antibody library screening (millions of candidates) or extensive fine-tuning on close analogs of the target. Chai-2 generates candidates directly against a previously unseen target structure, which compresses the upstream design phase from weeks or months to minutes.

The second is the experimental hit rate distribution. Chai Discovery reports that across 52 distinct antigen targets, the median experimental hit rate per target was 16 to 20 percent at the screening of 20 candidates per target. The previous baseline for purely computational antibody design was a fraction of one percent in published benchmarks, so the reported uplift is approximately two orders of magnitude. The hit rate metric is the fraction of designed candidates that show measurable binding affinity in laboratory testing.

The third is multi-target coverage. The 52 antigens included in the released benchmark span multiple difficulty classes, including some classified in the November 2025 update as "challenging" targets that the prior antibody design literature had not addressed. The breadth of target coverage at high hit rate suggests the model generalizes across antigen classes rather than overfitting to a narrow training distribution.

The architectural details are described in the bioRxiv preprints but are less fully characterized in public material than Chai-1's architecture, reflecting the closed source posture.

Benchmarks and standing

The principal disclosed benchmarks for Chai-2 are experimental hit rates rather than computational metrics.

On the 52-target zero-shot antibody design benchmark, Chai Discovery reported a 16 percent median hit rate per target across 20 candidates per target. At least one successful binder was identified for 50 percent of targets in a single round of experimental testing. Hit rates ranged up to approximately 20 percent on some targets. The November 2025 update extended the benchmark to "challenging" target classes including those characterized in coverage as previously unaddressed by computational antibody design.

In comparative terms, the prior computational antibody design baseline in published research had been reported at well under one percent experimental hit rates, so the Chai-2 result represents a 100 to 200 fold uplift on the headline metric. Industry coverage and Chai Discovery's own framing characterize this uplift as the principal capability advancement that separates Chai-2 from prior generative antibody design research.

Reproducibility is a structural open question. Independent third-party reproduction would require replicating the laboratory expression, characterization, and binding affinity measurement pipeline alongside the model. The Eli Lilly partnership gives Lilly the contractual position to verify the capability claims in production-style biologics discovery workflows.

The horizontal language model benchmarks (Artificial Analysis Intelligence Index, LMArena, GPQA Diamond, AIME, SWE-bench) do not apply to generative antibody design. The relevant benchmarks are zero-shot hit rates at standardized candidate counts, target diversity coverage, and downstream developability metrics (manufacturability, immunogenicity risk, half-life), some of which are explicitly addressed in the November 2025 paper.

Access and pricing

Chai-2 is not available through a hosted public web tool, an API, or open weights distribution. The model is closed source, and access is structured through commercial partnerships rather than self-service developer access.

The most prominent disclosed access channel is the multi-year Eli Lilly collaboration announced January 2026, under which Chai Discovery is developing a custom Chai variant trained on Lilly proprietary data and tailored to Lilly biologics discovery workflows. The collaboration covers multiple targets across Lilly biologics portfolio, with terms not publicly disclosed.

Chai Discovery also operates internal molecule design programs, applying Chai-2 to candidates that the company can advance through its own discovery pipeline either alone or in partnership.

Per token, per query, or per design pricing has not been publicly disclosed for Chai-2. The pharmaceutical industry's standard partnership economics (research access fees, milestone payments, and royalty structures on resulting therapeutics) are the implied commercial model, with terms negotiated bilaterally with each partner.

Comparison

Direct competitors and adjacent generative antibody design systems:

• AbCellera, BigHat Biosciences, Profluent. Commercial peers in generative antibody design. AbCellera and BigHat have published less aggressive zero-shot performance numbers than Chai-2 but operate established pharmaceutical partnership programs. Profluent has emphasized broader generative protein design rather than antibody design specifically. The published Chai-2 hit rates lead the disclosed peer benchmarks at the May 2026 reading, while reproducibility remains an open question.
• AlphaFold 3 and Isomorphic Labs antibody work (Isomorphic Labs, Google DeepMind). Isomorphic Labs has not published a directly comparable zero-shot antibody design hit rate benchmark, but its broader drug discovery pipeline runs against the same target classes Chai-2 addresses. The two companies are widely characterized as the leading commercial AI-for-biology efforts, with Chai's published hit rate evidence and Isomorphic Labs' Alphabet scale resources as the principal differentiators.
• EvolutionaryScale ESM3 (EvolutionaryScale). ESM3 spans broader biomolecular reasoning rather than focusing on antibody design specifically. ESMFold and ESM-based generative protein design lines provide alternative platforms, with antibody design as a downstream application rather than the central thesis.
• Generate Biomedicines, Recursion, Insitro. Larger and longer running AI drug discovery companies with portfolio approaches to therapeutic design. Generate Biomedicines has published generative antibody design work, while Recursion and Insitro emphasize phenotypic screening and target identification at scale rather than de novo antibody generation specifically.
• Chai-1 (Chai Discovery). The same lab's prior open source structure prediction model. Chai-1 and Chai-2 address structurally distinct tasks: Chai-1 predicts structures from sequences, while Chai-2 generates novel antibody designs against target structures. The two models complement each other in a complete antibody discovery pipeline.

Chai-2's distinctive position among 2025 vintage generative antibody design systems: the published 16 to 20 percent zero-shot hit rate at scale across 52 targets, the closed source posture that protects pharmaceutical partnership value, the Eli Lilly collaboration as a major-pharma reference customer, and the underlying Chai-1 structural prediction lineage that anchors research credibility.

Outlook

Open questions for Chai-2 over the next 6 to 18 months:

• Reproducibility evidence. The 16 to 20 percent hit rate evidence is currently a Chai Discovery published claim, with limited independent reproduction. Lilly partnership outputs, additional partner disclosures, or community reproduction of the published benchmark on representative target classes would clarify the result's reliability.
• Successor model cadence. Chai-2 was published in July 2025 with a November 2025 update. Whether Chai Discovery releases a Chai-3 or comparable successor in 2026 will signal the model line's trajectory and the durability of the published capability lead.
• Eli Lilly collaboration milestones. The January 2026 partnership covers multiple Lilly biologics targets. Disclosed milestones, including any Lilly clinical-stage candidate originating from Chai-2 designs, would be the strongest external validation.
• Pharmaceutical partnership expansion. The Lilly partnership is the most prominent disclosed Chai-2 commercial deployment. Whether the model lands additional named partner relationships through 2026 and 2027 will indicate the model's commercial traction at scale.
• Open source release possibility. Chai-1 was open source, and Chai-2 is closed. Whether Chai Discovery returns to a partial open source release at any point, including a research benchmark variant of Chai-2 or a specific design subtask, will affect the company's research community standing.

Sources

• Chai-2: Zero-shot antibody design in a 24-well plate (bioRxiv, July 2025). Technical preprint with the headline benchmark results.
• Drug-like antibody design against challenging targets (bioRxiv, November 2025). Updated benchmark on harder target classes.
• Chai Discovery blog: Introducing Chai-2. Official announcement.
• BioPharmaTrend: Chai-2 zero-shot antibody design. Release coverage.
• MarkTechPost: Chai-2 achieves 16% hit rate. Capability summary.
• Chai Discovery Eli Lilly collaboration announcement. January 2026 partnership.
• TechCrunch: Chai Discovery raises $130M Series B. December 2025 funding context.