Microsoft’s BioGPT recorded 86,100 monthly downloads on Hugging Face by early 2026, ranking it among the most-used domain-specific biomedical language models. The model holds 263 likes on Hugging Face and 4,500+ GitHub stars. This article covers BioGPT adoption numbers, benchmark performance, training scale, and where the biomedical AI market is heading.
Key BioGPT Statistics 2026
- BioGPT logs 86,100 monthly downloads on Hugging Face as of February 2026.
- The base model holds 347 million parameters and was trained on 15 million PubMed abstracts.
- BioGPT-Large reached 81% accuracy on the PubMedQA benchmark, beating prior records.
- 129 BioGPT-tagged models exist on Hugging Face Hub, including community fine-tunes.
- The global AI in healthcare market is projected at $54.19 billion in 2026, with generative AI growing at a 45.76% CAGR through 2031.
What Is BioGPT and Why Does It Matter?
BioGPT is a generative transformer language model built by Microsoft Research and released in 2022. The model holds 349 million parameters and is based on GPT-2 medium, with training focused only on biomedical text rather than general web data.
The model was published in the journal Briefings in Bioinformatics and made available through GitHub and Hugging Face. It is widely cited as one of the first GPT-style models tuned for medical literature mining and question answering.
BioGPT Download and Usage Statistics
BioGPT continues to draw heavy traffic on Hugging Face years after its release. The base model alone records tens of thousands of monthly downloads, and the wider BioGPT family covers 129 model variants on the platform.
| BioGPT Variant | Hugging Face Likes | Status |
|---|---|---|
| microsoft/biogpt (base, 347M) | 263 | Active |
| microsoft/BioGPT-Large | 197 | Active |
| microsoft/BioGPT-Large-PubMedQA | 114 | Active |
| Community fine-tunes (combined) | 290+ | Growing |
| Total BioGPT-tagged models | 129 | Indexed |
Source: Hugging Face Hub model registry
How Big Is the BioGPT Model?
BioGPT comes in two main sizes. The base model uses 347 million parameters, while BioGPT-Large reaches a 1.5 billion parameter scale through scaling on top of the GPT-2 XL base.
The base model was trained on 15 million PubMed abstracts pulled from publications dating back decades. Average abstract length came in at roughly 200 tokens.
| Specification | BioGPT (Base) | BioGPT-Large |
|---|---|---|
| Parameters | 347 million | 1.5 billion |
| Architecture base | GPT-2 medium | GPT-2 XL |
| Training corpus | 15M PubMed abstracts | 15M PubMed abstracts |
| Layers | 24 | 48 |
| Hidden size | 1,024 | 1,600 |
Source: Microsoft Research, Briefings in Bioinformatics
BioGPT Benchmark Performance
BioGPT set new state-of-the-art records on four benchmarks in its release paper. On the BC5CDR, KD-DTI, and DDI end-to-end relation extraction tasks, BioGPT scored 44.98%, 38.42%, and 40.76% F1 respectively, with 78.2% accuracy on PubMedQA. The larger BioGPT-Large model pushed PubMedQA accuracy to 81.0%.
| Benchmark | BioGPT Score | Task Type |
|---|---|---|
| BC5CDR | 44.98% F1 | Chemical-disease relation extraction |
| KD-DTI | 38.42% F1 | Drug-target interaction |
| DDI | 40.76% F1 | Drug-drug interaction |
| PubMedQA (BioGPT) | 78.2% accuracy | Yes/no/maybe question answering |
| PubMedQA (BioGPT-Large) | 81.0% accuracy | Yes/no/maybe question answering |
Source: Luo et al., Briefings in Bioinformatics, 2022
BioGPT vs ChatGPT on Biomedical Tasks
A 2023 ACL evaluation tested how BioGPT compares with ChatGPT on the same biomedical datasets. BioGPT scored 78.20% accuracy on PubMedQA against ChatGPT’s 51.60%, an absolute gap of 26.6 points. The domain-specific training pays off on tightly scoped medical tasks.
| Model | PubMedQA Accuracy | HoC F1 | DDI F1 |
|---|---|---|---|
| BioGPT | 78.20% | 85.12% | 40.76% |
| ChatGPT (zero-shot) | 51.60% | 59.14% | 46.43% |
Source: BioNLP Workshop 2023, ACL Anthology
How BioGPT Was Trained
Microsoft researchers used 15 million PubMed abstracts as the pre-training corpus. The team pulled articles published before 2021 from PubMed’s English-language biomedical database. Training ran for 200,000 steps using 8 NVIDIA V100 GPUs with mixed-precision and gradient accumulation.
The training used causal language modeling with a peak learning rate of 2 × 10⁻⁴ and inverse square root decay scheduling with 20,000 warm-up steps.
BioGPT Statistics by Use Case
Researchers have applied BioGPT to drug discovery literature mining, clinical documentation drafting, biomedical question answering, and named entity recognition. The Hugging Face library exposes specialized heads for token classification and sequence classification beyond the base text generation pipeline.
| Application | Task | Typical Output |
|---|---|---|
| Relation extraction | BC5CDR, DDI, KD-DTI | Entity-relation tuples |
| Question answering | PubMedQA | Yes/no/maybe label |
| Document classification | HoC (Hallmarks of Cancer) | Topic labels |
| Text generation | Term description, summarization | Free-form medical text |
| Named entity recognition | BioGptForTokenClassification | Tagged entities |
Source: Hugging Face Transformers documentation
Biomedical AI Market Context for BioGPT
BioGPT operates inside a fast-growing biomedical AI sector. The AI in healthcare market is estimated at $54.19 billion in 2026, up from $39.92 billion in 2025, with projections of $249.72 billion by 2031 at a 35.74% CAGR.
Generative AI inside healthcare is forecast to expand at a 45.76% CAGR between 2026 and 2031, the fastest of any segment. Domain-specific models like BioGPT sit at the heart of that segment.
| Year | AI in Healthcare Market | Generative AI in Healthcare |
|---|---|---|
| 2025 | $39.92B | $3.3B |
| 2026 | $54.19B | $4.7B |
| 2027 | ~$73.5B | $5.95B |
| 2030 | ~$181B | $11.9B |
| 2031 | $249.72B | ~$15.1B |
Source: Mordor Intelligence and Roots Analysis, 2026
BioGPT Adoption in Research Communities
Researchers have published more than 60 fine-tuned BioGPT derivatives on Hugging Face, covering tasks from clinical NER to specialized question answering. The original GitHub repository attracts thousands of stars and is mirrored across pharmaceutical and academic teams.
Outside Microsoft’s own checkpoints, third-party variants such as healthcare-focused fine-tunes appear on text generation, token classification, and question answering leaderboards. The model competes with PubMedBERT, BioBERT, and SciFive in the biomedical NLP stack.
Limitations Behind the BioGPT Numbers
BioGPT’s training data ends in 2021, which means it does not cover newer drug approvals, recent COVID-19 variants, or papers published after that cutoff. Absolute F1 scores of 38–45% on relation extraction also indicate room for improvement on harder extraction problems.
The model has no built-in safety layer for clinical use, and Microsoft has stated it is intended for research rather than diagnostic deployment. Newer biomedical foundation models trained on multimodal data are starting to compete with BioGPT on several benchmarks, as shown in comparative AI diagnostic studies.
BioGPT Compared to Other Biomedical Language Models
BioGPT is the largest open-weight generative biomedical model in widespread use. PubMedBERT and BioBERT are bigger by community footprint, but they are encoder-only and cannot generate text. The table below shows how the main biomedical NLP models stack up.
| Model | Type | Parameters | Generation Capable |
|---|---|---|---|
| BioGPT | Decoder (GPT) | 347M | Yes |
| BioGPT-Large | Decoder (GPT) | 1.5B | Yes |
| PubMedBERT | Encoder (BERT) | 110M | No |
| BioBERT | Encoder (BERT) | 110M | No |
| SciFive | Encoder-decoder (T5) | 770M | Yes |
Source: Hugging Face model cards and original publications
FAQs
What is BioGPT used for?
BioGPT handles biomedical text generation, relation extraction, question answering, and document classification. Researchers apply it to PubMed literature mining, drug-drug interaction extraction, and biomedical NER tasks where domain-specific vocabulary matters.
How accurate is BioGPT on PubMedQA?
BioGPT reached 78.2% accuracy on PubMedQA, while BioGPT-Large pushed that figure to 81.0%. Both numbers set new records at release and remain competitive with much larger general-purpose models on biomedical question answering.
How many parameters does BioGPT have?
The base BioGPT model has 347 million parameters across 24 transformer layers with 1,024 hidden units. BioGPT-Large scales up to 1.5 billion parameters using 48 layers and 1,600 hidden units, both based on the GPT-2 architecture.
Is BioGPT free to use?
Yes. BioGPT is open-source under an MIT license and available through Hugging Face Hub and the Microsoft GitHub repository. Researchers can download weights, fine-tune the model, and deploy it for non-clinical research without licensing fees.
Who created BioGPT?
Microsoft Research created BioGPT, with lead authors Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. The paper was published in the journal Briefings in Bioinformatics in 2022.
Sources
https://huggingface.co/microsoft/biogpt
https://academic.oup.com/bib/article/23/6/bbac409/6713511
https://github.com/microsoft/BioGPT
https://www.mordorintelligence.com/industry-reports/artificial-intelligence-in-healthcare-market
