Compare · 24 models
Biomedical language models you can run online
Biomedical language models extract entities from clinical text, query scientific publications, and reason over biological data such as proteins, molecules, and medical imaging for researchers and clinicians.
What biomedical language models models do
Biomedical natural language processing spans several specialised tasks, including extracting clinical entities from unstructured records, retrieving and synthesising literature, and reasoning over complex biological relationships. Standard general-purpose language models often struggle with high-density medical terminology, chemical notation, genomic identifiers, and irregular clinical syntax. The models in this category address these challenges through domain-specific training on corpora such as PubMed abstracts, full-text scientific articles, and de-identified clinical records.
Architectures in this group range from compact token-classification encoders engineered for high-throughput named entity recognition to large multi-billion-parameter generative models. Several systems incorporate multimodal components, allowing them to connect textual prompts directly to protein representations, nucleotide sequences, or diagnostic imaging modalities like CT and MRI scans. Others employ retrieval-augmented workflows or structured chain-of-thought reasoning to ground scientific claims and explain intermediate deductions.
Compared
- PaperQA2 (Crow)FutureHouseGrounded retrieval-augmented generation agent that answers scientific queries using cited literature.Commercial OKInstant
- BioMedGPT-R1PharMolix / OpenBioMedA biomedical language model providing chain-of-thought reasoning over text, small molecules, and protein sequences.Commercial OKReady
- BioReason-Pro RLwanglabA reinforcement-learning-tuned language model for biological reasoning and protein function prediction.Check termsOn demand
- BioReason-Pro SFTwanglabMultimodal reasoning model for protein function prediction with structured chain-of-thought rationales.Check termsOn demand
- ChatNTInstaDeepAIConversational model answering natural-language queries for DNA, RNA, and protein sequence tasks.Check termsOn demand
- DoctoBERT-fr-basedoctolib-labA 111M-parameter French RoBERTa encoder pretrained for biomedical and clinical text processing.Check termsOn demand
- DoctoModernBERT-fr-basedoctolib-labA 149M-parameter French biomedical encoder with an 8,192-token context window for clinical NLP and token classification.Check termsOn demand
- OpenMed AnatomyDetectOpenMedA 109M-parameter transformer for detecting anatomical structures, body parts, and tissues in biomedical text.Check termsOn demand
- OpenMed BloodCancerDetectOpenMedA compact 65M-parameter model that identifies chronic lymphocytic leukaemia clinical entities in biomedical text.Check termsOn demand
- OpenMed ChemicalDetectOpenMedA 149M-parameter transformer for detecting chemical compounds and drug names in biomedical and clinical text.Check termsOn demand
- OpenMed DiseaseDetectOpenMedA biomedical transformer for detecting disease names and medical conditions in scientific and clinical text.Check termsOn demand
- OpenMed DNADetectOpenMedA 125M-parameter RoBERTa model for extracting DNA, RNA, protein, cell line, and cell type entities from biomedical text.Check termsOn demand
- OpenMed GenomeDetectOpenMedA 149M-parameter ModernBERT model for extracting gene and protein entity mentions from English biomedical text.Check termsOn demand
- OpenMed GenomicDetectOpenMedIdentifies and extracts genomic and cell-line entity mentions from biomedical literature and clinical texts.Check termsOn demand
- OpenMed OncologyDetectOpenMedA biomedical transformer for named entity recognition of cancer genetics and anatomical entities in text.Check termsOn demand
- OpenMed OrganismDetectOpenMedA compact biomedical transformer model for detecting species and organism mentions in scientific literature.Check termsOn demand
- OpenMed PharmaDetectOpenMedA transformer model for identifying and extracting chemical and drug entities from biomedical and clinical text.Check termsOn demand
- OpenMed ProteinDetectOpenMedA 141M-parameter DeBERTa model that identifies proteins, complexes, and variants in biomedical literature and clinical text.Check termsOn demand
- OpenMed SpeciesDetectOpenMedA 109M-parameter transformer for detecting species and organism names in biomedical literature and clinical texts.Check termsOn demand
- Clinical Camel 70BwanglabA 70-billion-parameter clinical language model fine-tuned for medical dialogue and clinical decision support.See its pageNot yet
- MedASRgoogleA 105-million-parameter speech recognition model fine-tuned on clinical dictation for medical transcription.See its pageNot yet
- MedGemma 1.5 4BgoogleA 4-billion-parameter multimodal model interpreting medical imaging and clinical text for biomedical researchers.See its pageNot yet
- TxGemma 27B ChatgoogleA 27-billion-parameter conversational language model for therapeutic discovery and drug property prediction.See its pageNot yet
- TxGemma 9B ChatgoogleA conversational 9B-parameter language model for drug discovery and therapeutic property prediction with reasoning.See its pageNot yet
How to choose
For literature querying and structured dialogue, the choice depends on evidence grounding and domain breadth. PaperQA2 is structured specifically for literature synthesis, using a three-phase search, evidence-gathering, and answer generation algorithm that ties every statement to retrieved passages and verifies retractions. If dialogue across drug discovery and therapeutic properties is required, TxGemma 9B Chat and TxGemma 27B Chat provide multi-turn conversational reasoning over small molecules, proteins, nucleic acids, and diseases. For step-by-step reasoning across small molecules and proteins, BioMedGPT-R1 outputs explicit chain-of-thought traces, though it is limited to a 2,048-token context length and requires 40 GB of GPU memory. Clinical Camel 70B uses a 70-billion-parameter foundation tuned on MedQA reasoning chains for clinical conversation, but demands substantial computational hardware.
For tasks bridging biological sequences or non-text modalities, specialised multimodal models are needed. ChatNT processes up to approximately 12 kilobases of DNA, RNA, or protein sequences across 27 biological instruction tasks using an 8-billion-parameter architecture that requires 40 GB of GPU memory. When predicting protein function and variant impact, BioReason-Pro SFT incorporates ESM3 embeddings, Gene Ontology graphs, and STRING networks into a 4-billion-parameter backbone, while BioReason-Pro RL applies reinforcement learning via Group Relative Policy Optimization for sharper deductions across KEGG pathways. For multimodal imaging, MedGemma 1.5 4B pairs a SigLIP image encoder with a decoder backbone to interpret 2D clinical photographs, digital histopathology, and 3D CT or MRI scans. For medical speech transcription, MedASR provides a dedicated 105M-parameter Conformer model optimised for clinical dictation waveforms.
For entity extraction and token classification in English text, the OpenMed family provides dedicated single-purpose encoders. For chemical and drug entities, OpenMed PharmaDetect (434M parameters) achieves 0.9614 F1 on BC5CDR-Chem, whereas OpenMed ChemicalDetect (149M parameters) provides a lighter alternative achieving 0.92 F1 on BC4CHEMD. OpenMed DiseaseDetect targets medical conditions, OpenMed OncologyDetect classifies 16 entity types across cancer genetics and pathology, OpenMed ProteinDetect extracts protein complexes and variants, and OpenMed BloodCancerDetect delivers a compact 65M-parameter model focused on chronic lymphocytic leukaemia mentions. For French clinical text, DoctoBERT-fr-base offers a 111M-parameter encoder with a 512-token context window, while DoctoModernBERT-fr-base provides a 149M-parameter encoder supporting sequences up to 8,192 tokens natively.
Questions
Can I run these models without a GPU or installing anything?
19 of the 24 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
Which models can process molecular sequences alongside natural language instructions?
ChatNT processes DNA, RNA, and protein sequences up to roughly 12 kilobases across 27 biological tasks. BioMedGPT-R1 reasons over natural text, small-molecule data, and protein sequences using chain-of-thought deduction. BioReason-Pro SFT combines text with ESM3 protein embeddings and biological interaction graphs, while TxGemma 9B Chat and TxGemma 27B Chat accept SMILES, amino acid, and nucleotide inputs for therapeutic property prediction.
How do DoctoBERT-fr-base and DoctoModernBERT-fr-base differ for French medical text processing?
Both models are non-generative encoders pretrained on the FineMed-fr corpus. DoctoBERT-fr-base is a 111-million-parameter RoBERTa model with a 512-token context window. DoctoModernBERT-fr-base has 149 million parameters, uses the ModernBERT architecture, supports contexts up to 8,192 tokens, and is compatible with FlashAttention-2.
How does PaperQA2 avoid fabricating citations in scientific writing?
PaperQA2 follows a three-phase algorithm: Paper Search, Gather Evidence, and Generate Answer. It retrieves candidate documents from external sources such as Europe PMC or local files, gathers and verifies metadata via providers like Crossref and Semantic Scholar, carries out retraction checks, and ensures every generated claim is tied directly to retrieved passages.
What is the difference between BioReason-Pro SFT and BioReason-Pro RL?
BioReason-Pro SFT is a supervised fine-tuned model combining a Qwen3-4B backbone with ESM3 protein embeddings, Gene Ontology graph encoders, and STRING interactions for multimodal protein function prediction. BioReason-Pro RL is an optimised checkpoint built on the SFT base using Group Relative Policy Optimization (GRPO) to generate higher-quality biological reasoning chains across KEGG pathways and variant datasets.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.