Proprietary 41B-parameter mixture-of-experts transformer for drug discovery, predicting 6,000+ molecular, preclinical and clinical properties.
No providers recorded yet. Browse all providers
A medicinal chemist optimizing a lead compound juggles dozens of properties at once: potency against the intended target, selectivity against its close relatives, metabolic stability, permeability, brain penetration and, ultimately, how the molecule behaves in patients. Each of those endpoints is usually modeled separately, from whatever small dataset exists for it, and the clinical endpoints that matter most have the least data. Enchant v3 takes the opposite approach. It trains one transformer on thousands of endpoints and many kinds of biomedical data together, so that a sparsely measured property can borrow signal from related assays, related molecules and other modalities.
Enchant v3 was developed by Iambic Therapeutics in San Diego and announced on 21 September 2026 in a press release and a technical report by Fred Manby and Matt Welborn. It is the third generation of the Enchant family, following Enchant v1 (1 billion parameters, trained in 2024) and Enchant v2 (7 billion parameters, trained in 2025). Iambic describes v3 as an architectural change rather than only a larger model: it moves to a mixture-of-experts transformer, in which each input activates a subset of specialized sub-networks rather than the whole model.
Enchant v3 is an internal platform model. Iambic deploys it across its own pipeline and partner programs and has released no code, weights, API or peer-reviewed paper.
Enchant v3 is a multimodal transformer with 41 billion parameters in a mixture-of-experts configuration, pretrained on 4.5 trillion tokens drawn from public, partnered and proprietary data. In Iambic's scaling analysis, a composite benchmark score rises log-linearly with training compute across the three generations, from about 60.7 for Enchant v1 to about 63.8 for Enchant v2. Enchant v3 scores about 66.2 at roughly 2×10²³ FLOP, the same score the dense-model trend reaches only near 10²⁴ FLOP. Iambic has not published the composition of the composite benchmark, the number of experts or the routing scheme.
Protein structures are one of the training modalities, included to improve predictions for assays that involve specific proteins. As a by-product, Enchant has learned to predict folded protein structure as a token-based method. On held-out structures it reached a TM-score of 0.775 on CAMEO22 and 0.575 on CASP14, against 0.746 and 0.534 for ESM-3 and 0.694 and 0.500 for DPLM-2. All reported results come from the developer and have not been independently reproduced.
Iambic uses Enchant to score every compound considered in its discovery programs against a target compound profile, fine-tuning the model each week on hundreds of molecular properties and running about 12 million inferences per month. The report credits Enchant with two development candidates. IAM-C1 is a dual CDK2/CDK4 inhibitor designed to spare closely related kinases such as CDK6, and it hits only four off-targets in a KINOMEscan panel. IAM217 is an allosteric KIF18A inhibitor for triple-negative breast, ovarian and other solid tumors, which produced about 90 percent regression of intracranial tumors in preclinical studies. Both are advancing toward IND submission. Iambic also applies Enchant v3 in pharmaceutical partnerships.
Across three generations, the Enchant family extends Iambic's claim that transformer scaling laws hold for drug-discovery data. Enchant v3's move to a mixture-of-experts design tests whether sparse architectures deliver the same compute savings on heterogeneous biomedical data that they do in language modeling. Outside researchers can only weigh these claims through Iambic's reports. The model is proprietary, its benchmarks are undisclosed composites and internal assays, and its value will ultimately be judged by whether the compounds it helped design succeed in the clinic. NeuralPLexer, Iambic's structure-prediction line, complements it within the same platform.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.