Models

Ensembles, not single structures — essential for disordered and dynamic proteins.

Fetch a pre-computed structure for any UniProt accession — 247M entries, instantly.

The reference MSA-based predictor, with managed MSA search included.

Structure and binding affinity in one pass, approaching FEP accuracy 1000× faster.

Open multimodal complex prediction across proteins, nucleic acids and ligands.

MMseqs2-accelerated AlphaFold2 — the fastest route to an MSA-quality fold.

High-quality 1,280-dimensional protein embeddings for everything downstream.

The throughput workhorse of the ESM2 family — good embeddings, tiny footprint.

300M-parameter ESMC protein language model — compact variant of the ESMC family for protein-embedding workflows, fits on a single consumer GPU.

600M-parameter ESMC protein language model — provides digital representations of proteins for therapeutic engineering, variant-effect prediction, and transfer learning across protein tasks.

Flagship 6B-parameter ESMC protein language model — state-of-the-art protein embeddings for variant-effect prediction, protein engineering, and biology research, trained on billions of protein sequences.

Single-sequence folding in seconds — no MSA, no waiting.

State-of-the-art protein structure prediction and design model — defines a new frontier for speed and accuracy, with unprecedented success rates for protein binder and antibody generation.

Experimental ESMFold2 variant exploring new training recipes and architectural changes for protein structure prediction, design, and confidence estimation.

Fast variant of the experimental ESMFold2 line — combines accelerated inference with the research-recipe improvements of the experimental branch.

Accelerated variant of ESMFold2 — optimised for high-throughput structure prediction across large protein sets while retaining strong accuracy.

Open-weight complex prediction across proteins, nucleic acids and ligands, Apache-2.0 throughout.

MSA-free folding tuned for orphan sequences with no homologs.

A fully retrained, Apache-2.0 reproduction of AlphaFold2.

Experimentally determined coordinates and metadata for any PDB entry.

Full-atom prediction including non-canonical residues and cofactors.

Composition, molecular weight, isoelectric point and hydropathy — deterministic.

AF2-guided hallucination that co-folds target and binder — ~10 screens to a hit.

Diffusion protein language model generating sequence and structure jointly.

Language-model-informed inverse folding — an alternative view to ProteinMPNN.

Few-shot active learning for directed evolution — 24 mutants per round.

Inverse folding that sees the ligand: sequence design conditioned on the atoms around the protein.

Autoregressive protein generation up to 6.4B parameters, with an antibody variant.

Inverse folding: given a backbone, design sequences that actually fold onto it.

De novo backbone generation by denoising diffusion — the binder-design workhorse.

Structure-aware protein LM using combined sequence + structure tokens.

The docking engine every other docking paper compares itself against.

Diffusion docking: 38% top-1 success versus 23% for classical AutoDock.

Diffusion docking for protein–protein interfaces.

CNN rescoring on top of Vina poses — usually a real enrichment gain.

End-point free energy — faster than FEP, good for ranking close analogues.

Full FEP network setup and execution when estimates need experimental grade.

A Vina fork with custom scoring — the fragment-docking specialist.

Antibody-specific language model trained on the Observed Antibody Space.

Antibody numbering and region annotation — Chothia, IMGT, Kabat.

Inverse folding specialised for antibody backbones and CDR loops.

Antibody structure prediction ~1000× faster than AlphaFold2.

Sub-second structure prediction for antibodies, nanobodies and TCRs.

Binary protein-solubility classification from amino-acid sequence.

Paratope and epitope prediction for humanization and immunogenicity work.

Germline-guided antibody humanization, validated on clinical-stage molecules.

Find the protein-coding genes in a eukaryotic genome: exons, strand, reading frame and a confidence per gene.

Ask a question about DNA in plain English, or rank variants by how typical the model finds them.

100M-cell foundation model for pan-tissue annotation and cross-dataset integration.

BPE-tokenised genomic BERT for motif, splice site and regulatory classification.

Sequence-to-expression across a 200kb window covering distal enhancers.

Genome-scale foundation model spanning nucleotide to whole-genome context.

Transformer over 30M transcriptomes for network inference and target prioritisation.

Generate an RNA sequence and its 3D structure together, from a length — unconditional, so no target fold or function.

DNA foundation model for regulatory elements, promoters and variant scoring.

Gene regulatory network inference from paired single-cell ATAC and RNA.

Single-cell foundation model trained on 33 million human cells.

52 ADMET endpoints in one call — hERG, BBB, CYP, clearance, solubility, Tox21.

AIMNet2 retrained at the B97-3c level of theory — 2025 release with improved coverage and accuracy.

Neural network interatomic potential for fast and accurate molecular simulations, trained at the ωB97M-D3 level of theory.

AIMNet2 specialised for open-shell chemistry (radicals, transition states) — neural network interatomic potential for non-singlet electronic states.

AIMNet2 specialised for palladium-containing organometallic systems — supports homogeneous catalysis simulation at near-DFT accuracy.

AIMNet2 trained on reaction data — neural-network interatomic potential supporting reactive molecular simulations.

The toxicity half of the ADMET panel: Ames, hERG, DILI, ClinTox, LD50 and all twelve Tox21 assays.

Joins 3D fragments into one molecule by generating the atoms in between.

Equivariant diffusion that generates molecules inside a given binding pocket.

Self-supervised molecular GNN, strong across ADMET endpoints.

MACE foundation model trained on the Materials Project — equivariant message-passing potential for inorganic crystal simulation across most of the periodic table.

MACE foundation model trained on the Materials Project + Alexandria datasets — broader coverage variant for inorganic-materials simulation.

Autoregressive SMILES generation and forward reaction prediction.

Lipinski, Veber and lead-likeness descriptors computed from SMILES — instantly, free.

SMILES transformer pre-trained on 1.1 billion molecules.

Controlled molecule generation in a well-behaved latent space.

Forward and retrosynthetic prediction, plus RXNFP fingerprints for comparing reactions.

RL-guided generative chemistry against a multi-objective scoring function.

Frontier one-step retrosynthesis: what would you make this from, ranked by an ensemble of two models that disagree usefully.

Design molecules that present the same surface to a binding site, without keeping the scaffold.

One model pretrained across molecules, materials and proteins; here it draws 3D molecules and predicts their energy.

First 1.0 release of the deterministic AIFS — graph-based global weather forecasting model from ECMWF on the Anemoi framework.

OpenCLIP-based foundation model for organismal biology — zero-shot species classification from photographs across the tree of life, trained on TreeOfLife-200M.

Reasoning-capable biomedical LLM with chain-of-thought over molecules.

MAMMAL embeddings bridged into a Granite LLM for zero-shot cross-modal Q&A.

French medical encoder (RoBERTa, 111M params, 512-token context) pretrained from scratch on FineMed-fr — state-of-the-art on DrBenchmark, leading five of seven tasks across French and English medical encoder families.

French medical encoder (ModernBERT, 149M params, up to 8,192-token context) pretrained from scratch on FineMed-fr — best on real-world clinical NER and top-ranked on DrBenchmark.

Fast protein-protein interaction prediction model — trained for high-throughput screening of interaction networks.

Scientific LLM retained as a domain-knowledge evaluation baseline only.

BERT-base-style genomic foundation model trained on T2T assemblies — lighter-weight backbone for genomic sequence understanding.

BERT-large-style genomic foundation model trained on telomere-to-telomere human assemblies — supports variant interpretation, regulatory prediction, and downstream genomic tasks.

650M-parameter genomic foundation model from Tatta Bio — trained on the OMG open-mixed-genomes corpus for sequence-level biological reasoning.

Science domain extraction model for identifying and classifying scientific concepts, variables, and entities from geoscience and atmospheric science text.

Task-specific predictions for binding affinity, blood–brain barrier permeability, and protein solubility.

MIST 1.8B fine-tuned for electronic spatial extent from QM9.

MIST 1.8B fine-tuned on BACE — Alzheimer-target inhibition classification.

MIST 1.8B fine-tuned on BBBP — blood-brain-barrier permeability.

MIST 1.8B fine-tuned on ClinTox — clinical toxicity classification.

MIST 1.8B fine-tuned for heat capacity Cv from QM9.

MIST 1.8B fine-tuned on ESOL — aqueous solubility regression.

MIST 1.8B fine-tuned on FreeSolv — hydration free-energy regression.

MIST 1.8B fine-tuned for G298 — Gibbs free energy at 298 K from QM9.

MIST 1.8B fine-tuned for H298 — enthalpy at 298 K from QM9.

MIST 1.8B fine-tuned on HIV — anti-HIV activity classification.

MIST 1.8B fine-tuned for HOMO energy from QM9.

MIST 1.8B fine-tuned for HOMO-LUMO gap from QM9.

MIST 1.8B fine-tuned on Lipophilicity — large-scale logD prediction.

MIST 1.8B fine-tuned for LUMO energy from QM9.

MIST 1.8B fine-tuned on QM8 — electronic-spectra prediction at scale.

MIST 1.8B fine-tuned on SIDER — side-effect prediction.

MIST 1.8B fine-tuned on Tox21 — large-scale toxicity classification across nuclear-receptor and stress assays.

MIST 1.8B fine-tuned for U0 — internal energy at 0 K from QM9.

MIST 1.8B fine-tuned for U298 — internal energy at 298 K from QM9.

MIST 1.8B fine-tuned for zero-point vibrational energy from QM9.

MIST 1.8B fine-tuned for isotropic polarizability from QM9.

MIST 1.8B fine-tuned for dipole moment from QM9.

MIST 1.8B base — large pretrained molecular language model (fill-mask) for downstream chemistry property prediction at scale.

MIST 26.9M fine-tuned for boiling-point regression.

MIST 26.9M fine-tuned for dn property regression.

MIST 26.9M fine-tuned for flash-point regression.

MIST 26.9M fine-tuned for melting-point regression.

MIST 26.9M fine-tuned for odour-quality prediction.

MIST 27.0M fine-tuned for ionic-conductivity prediction in chemical mixtures and electrolytes.

MIST 27.1M fine-tuned on the ETN (empirical thermodynamic network) benchmark.

MIST 28M fine-tuned on BACE — beta-secretase 1 (Alzheimer target) inhibition classification.

MIST 28M fine-tuned on BBBP — blood-brain-barrier permeability classification for CNS drug candidates.

MIST 28M fine-tuned on ClinTox — clinical toxicity classification of FDA-approved drugs and failed candidates.

MIST 28M fine-tuned on ESOL — aqueous solubility regression (Delaney dataset).

MIST 28M fine-tuned on FreeSolv — hydration free-energy regression for small molecules.

MIST 28M fine-tuned on HIV — anti-HIV activity classification from MoleculeNet.

MIST 28M fine-tuned on Lipophilicity — octanol/water distribution coefficient prediction.

MIST 28M fine-tuned on MUV — maximum-unbiased-validation virtual-screening benchmark.

MIST 28M fine-tuned for pKa — acid-dissociation-constant prediction.

MIST 28M fine-tuned on QM8 — electronic-spectra property prediction over small organic molecules.

MIST 28M fine-tuned on SIDER — side-effect prediction across 27 system-organ classes for marketed drugs.

MIST 28M fine-tuned for solvent-property prediction — bulk physical descriptors of organic solvents.

MIST 28M fine-tuned on tmQM — quantum-mechanical property prediction for transition-metal complexes.

MIST 28M fine-tuned on Tox21 — toxicity classification across 12 nuclear-receptor and stress-response assays.

MIST 28M fine-tuned on ToxCast — multi-task toxicity prediction across hundreds of in-vitro assays.

MIST 28M base — pretrained molecular language model (fill-mask) used as the starting point for downstream property-prediction fine-tunes.

MIST mixtures variant — pretrained on chemical mixtures rather than individual molecules.

Compact ModernBERT-based GENA-LM variant — efficient genomic foundation model for downstream variant and expression tasks.

GENA-LM rebuilt on the ModernBERT architecture — larger, longer-context, RoPE-equipped genomic foundation model.

RoBERTa-based language model pre-trained on NASA Science Mission Directorate literature for earth and space science information extraction.

2.5B-parameter DNA language model trained on 850 species genomes — state-of-the-art on promoter, enhancer, and splice site prediction tasks.

Masked-autoencoder foundation model for high-content cell imaging — learns phenomic embeddings from millions of microscopy images for downstream drug-discovery and perturbation analysis.

Qwen3-VL fine-tuned to generate parametric CAD models directly from images — bridges vision-language reasoning and engineering geometry synthesis.

The Amber force field family plus the full analysis toolchain.

Massively parallel Lamarckian genetic algorithm docking for library screens.

The high-performance MD workhorse for large systems and membranes.

GPU-native molecular dynamics with a scriptable Python API.

Simulate mitochondrial ATP synthesis, oxygen consumption and membrane potential.

Forecast a series it has never seen, with no fitting and no training on your data.

Llama-2 7B fine-tuned on 300K+ astronomy arXiv abstracts for astrophysics text generation, literature summarization, and hypothesis completion — first open LLM specialized for astronomy.

RL-tuned variant of BioReason-Pro — reinforcement-learning fine-tuning over BioReason’s SFT base for sharper biological reasoning across KEGG pathways and variant data.

Supervised fine-tuned variant of BioReason-Pro — multimodal biological reasoning over genomic variants and pathway data with chain-of-thought rationales.

Flagship 3B-parameter genomic foundation model from Hugging Face Biology Research — Llama-architecture DNA language model trained for sequence generation, variant interpretation, and downstream genomic tasks.

Compact 500M-parameter Carbon variant — DNA language model lightweight enough for downstream fine-tuning on consumer hardware.

Largest 8B-parameter Carbon variant — top-of-line DNA language model for the most demanding genomic prediction, generation, and benchmark tasks.

8B multimodal conversational model for DNA, RNA, and protein tasks — instruction-following for sequence annotation, classification, and generation.

24B parameter model for molecular reasoning - SMILES generation, property prediction, and retrosynthesis.

General-purpose 1.2B-parameter genomic foundation model spanning multiple organisms — base model for downstream gene-level and sequence-level prediction tasks.

Medical multimodal LLM from the HuatuoGPT family — answers clinical questions over medical imagery (radiology, pathology, dermatology) using a 7B vision-language backbone.

Compact 1.7B distilled Kimina-Prover variant for Lean 4 proof generation on olympiad-level theorems — runs on a single consumer GPU.

8B distilled Kimina-Prover variant — Lean 4 theorem-proving model trained on olympiad-level mathematical problems with reinforcement learning over proof traces.

Distilled 7B preview of Kimina-Prover — a reinforcement-learning-trained model that generates Lean 4 proofs for olympiad-level mathematics problems.

MarinDNA m5.1 is a 1.12B-parameter, nucleotide-level causal language model developed with Marin.

Mixtral-architecture genomic foundation model specialised for rice (Oryza sativa) — supports variant analysis, expression prediction, and breeding-relevant trait modelling.

Anatomical-entity NER for biomedical text — labels body parts, organ systems, and tissue references using the ElectraMed 109M backbone.

Compact 65M token-classification model that identifies haematologic malignancy mentions (leukaemia, lymphoma, myeloma subtypes) in clinical and biomedical text.

Chemical-entity NER over biomedical literature — identifies drug names, compounds, and chemical substances using the ModernMed 149M backbone.

Disease-mention NER trained on the BioMed 335M backbone — recognises disease names, syndromes, and condition references in clinical and biomedical literature.

DNA-mention NER for biomedical text — extracts gene-level DNA sequence references and locus identifiers using the SuperMedical 125M backbone.

Genome-mention NER complementary to GenomicDetect — focuses on whole-genome and assembly-level references in biomedical text.

Genomic-entity NER over PubMed-style text — labels genes, transcripts, and other genomic references for downstream knowledge extraction.

Oncology-focused NER that identifies cancer-type mentions, tumour grading, and staging language across clinical and biomedical literature.

Organism-mention NER for biomedical text — broader than SpeciesDetect, also picking up genera, strains, and informal organism references.

Pathology-finding NER over clinical and biomedical text — surfaces histopathological observations, lesion descriptions, and tissue-level abnormalities.

Token-classification model for pharmaceutical entity recognition in clinical text — built on the SuperClinical 434M backbone for high-recall drug, dose, and regimen extraction.

Protein-mention NER for biomedical and clinical text — extracts protein names, family references, and post-translational modification descriptors.

Species-mention NER over biomedical literature — identifies organisms and taxonomic references using the ElectraMed 109M backbone.

GPT-NeoX-based generative language model for antibody sequence design, trained on the Observed Antibody Space to generate diverse immunoglobulin heavy and light chains.

Grounded literature retrieval over the biomedical corpus — every claim cited.