Iambic launches Enchant v3, a 41B multimodal drug-discovery model

Iambic unveiled Enchant v3 — a 41-billion-parameter multimodal transformer pretrained on 4.5 trillion tokens across more than 6,000 molecular properties and 16 biomedical data types — as the next brain of its molecular-superintelligence platform for end-to-end drug discovery and development.

HEALTH desk — drug discovery is getting the same scaling playbook as language models: more parameters, more data types, better predictions. Enchant v3 is that bet written as a product launch.

Company specs on that page: Enchant v3 has 41 billion parameters — the adjustable weights inside the model, a rough size measure — and was pretrained on 4.5 trillion tokens, the small chunks of data the model sees while it learns. It is trained across more than 6,000 molecular properties and 16 biomedical modalities. Multimodal here means many kinds of data in one model, not text alone. File those 41-billion / 4.5-trillion / 6,000-plus / 16-modality lines as Iambic’s. This desk did not count the weights or the tokens.

The 16 modalities Iambic names include text, assay data, molecule structure, protein structure, 3D structure, images, multi-omics, biologics, clinical trial data, and pharmacokinetics — how a body absorbs and clears a drug. File that named list as Iambic’s. Do not invent extra modalities or a score versus AlphaFold or any other named rival. None was printed.

Architecture notes, still company: Enchant v3 uses a mixture-of-experts design — different specialist sub-networks inside one model — ingests public and proprietary data, and tests multiple hypotheses in parallel. It converts outputs into probability-guided decisions with uncertainty quantification, so the lab can prioritize which compounds to try next. File that mixture-of-experts / parallel-hypotheses / uncertainty-quantification picture as Iambic’s. This desk did not watch a screening run.

What Enchant was built to do, still Iambic: ingest diverse data; get better at predicting one endpoint when trained on related endpoints and other molecules; break through data walls that have historically separated preclinical discovery from clinical development; and turn model outputs into lab decisions. An endpoint here is a measured outcome the model is trying to guess — a lab assay, a safety mark, a clinical readout. File that ingest / transfer / data-wall / lab-decision picture as Iambic’s. This desk did not measure an endpoint.

Named voice on the release: Fred Manby, PhD, co-founder and CTO. He calls Enchant v3 “an architectural leap over Enchant v2, with new capabilities to be deployed for internal and partner drug discovery and development efforts.” He says earlier Enchant versions helped predict critical preclinical and clinical endpoints, that Enchant has shown robust scaling laws — larger models, stronger predictions — and that v3 is the next step “to make better technology for better medicines.” File the name, title, and those attributed lines as his, via Iambic. Do not upgrade “helped us predict” into a named clinical success, a trial win, or an approved drug. His other sentences stay in Sources.

Named voice, still the same page: Matt Welborn, PhD, SVP of Machine Learning. He says that when Enchant launched in 2024 it was not obvious scaling laws would hold in drug discovery; that biomedical data is richer than text and harder to work with; and that across three generations, whenever Iambic added parameters, data, and modalities, the predictions improved — including for endpoints where data is sparse. File the name, title, the 2024 launch line, and that scaling-held / sparse-endpoint line as his, via Iambic. Do not invent a head-to-head score versus a named rival.

Company objectives for v3: increase the probability of success for drug candidates, and expand into new therapeutic areas and modalities for more internal and partner programs. File those two aims as Iambic’s. This desk is not calling a candidate a success and is not inventing a partner roster, an open-weights release, or a price. None was printed on the launch post.

Same-day syndication: BusinessWire carried the same originating text at about 07:00 ET, as reprinted on FinancialContent. File that as the same-day company-wire carry — not a second originating newsroom, and not a source for invented benchmark wins. Extra about-box color — a novel drug candidate advanced to clinic in about two years, NeuralPLexer as a sister technology — stays in Sources. Do not read that about-box line as an Enchant v3 clinical win.

Plain English for the rest of the card: multimodal = many kinds of data in one model — text, assays, structures, images, omics, and more — not text alone. parameters = the adjustable weights inside the model; 41 billion is Iambic’s size figure. tokens = the small chunks of data the model sees while it learns; 4.5 trillion is Iambic’s pretraining count. mixture-of-experts = different specialist sub-networks inside one model. uncertainty quantification = the model also saying how sure it is, so the lab can pick the next experiment. endpoint = a measured outcome the model is trying to guess. pharmacokinetics = how a body absorbs and clears a drug. molecular superintelligence platform = Iambic’s name for Enchant plus its in-house chemistry and biology lab. This is a company model launch, not an open-weights drop, not a priced product, and not a new drug approval.

PRIMARY here: Iambic’s 21 Sep 2026 company post “Iambic Launches Enchant v3 – Molecular Superintelligence Designed to Advance End-to-End Drug Discovery & Development” — Tier A PRIMARY company source, the original record. The FinancialContent BusinessWire carry at about 07:00 ET is same-day syndication of that originating text — not a substitute primary, and not a source for invented scores versus named rivals. The 41-billion-parameter / 4.5-trillion-token / 6,000-plus-property / 16-modality specs, the named modality list, the mixture-of-experts / parallel-hypothesis / uncertainty-quantification picture, the ingest / data-wall / lab-decision job, the Manby and Welborn titles and attributed lines, the 2024 Enchant launch, the scaling-laws claim, and the two v3 objectives are company-attributed. Product behavior and any pipeline examples stay company-attributed — not independently audited here. NOT claimed: an open-weights release, a price, a head-to-head score versus AlphaFold or any other named rival, a named clinical-trial win beyond Iambic’s “helped us predict” line, a new drug approval, that this desk trained or tested Enchant v3, a stock tip, or investment advice. Distinct from the already-filed abbvie-iambic-ai-drug, novo-anthropic-claude-drug-discovery, inductive-bio-indy-chemistry, insilico-longevity-vaccines, and mithrl-20m-series-a.