FinBERT, the financial sentiment model originally released by Prosus AI, recorded over 3.5 million downloads on Hugging Face in the past month alone, making it one of the most-used domain-specific NLP models in finance. This article covers the latest adoption, accuracy, training data, and ecosystem statistics around FinBERT in 2026.
Key FinBERT AI Statistics 2026
- ProsusAI/finbert recorded 3,508,680 downloads on Hugging Face in the last 30 days.
- The model achieves 97% accuracy on the full-agreement subset of the Financial PhraseBank.
- FinBERT was trained on a 4.9 billion token financial communication corpus.
- 87 fine-tuned variants of ProsusAI/finbert are publicly hosted on Hugging Face.
- FinBERT2, released at KDD 2025, was pretrained on 32 billion finance-specific tokens.
How Many People Use FinBERT AI in 2026?
The original ProsusAI/finbert model recorded 3,508,680 downloads on Hugging Face in the last month, while the academic finbert-tone variant from yiyanghkust added another 947,386 downloads.
The Xenova ONNX port for browser deployment added 6,284 downloads, and the FinnSentiment fine-tune reached 11,439. Together, the top FinBERT family models exceed 4.4 million monthly downloads.
| FinBERT Variant | Monthly Downloads | Primary Use Case |
|---|---|---|
| ProsusAI/finbert | 3,508,680 | News sentiment classification |
| yiyanghkust/finbert-tone | 947,386 | Analyst report tone analysis |
| fergusq/finbert-finnsentiment | 11,439 | Finnish-language sentiment |
| Xenova/finbert (ONNX) | 6,284 | In-browser inference |
| Narsil/finbert | 36 | Mirror copy |
Source: Hugging Face Model Hub
FinBERT AI Statistics 2026 on Accuracy and Performance
On the Financial PhraseBank dataset where all annotators agreed on the label, FinBERT reached 97% accuracy, six percentage points above the prior state-of-the-art FinSSLX model. On the full dataset including disputed sentences, accuracy was 86%, a 15-point gain over the previous best.
| Model | Accuracy (Full Agreement Subset) | F1 Score |
|---|---|---|
| FinBERT | 97% | 0.95 |
| FinSSLX | 91% | 0.86 |
| HSC | 71% | 0.76 |
| LPS | 71% | 0.79 |
| Vanilla BERT | 96% | 0.94 |
Source: Araci, FinBERT paper, arXiv 1908.10063
Recent Benchmark Results in 2025
A November 2025 reproducible benchmark study found that zero-shot FinBERT scored a macro F1 of 0.555 on a sector-tagged financial news corpus. After fine-tuning on 1,500 manually labelled headlines, the same model reached macro F1 of 0.707.
On the cleaner Financial PhraseBank, a hybrid pairing FinBERT embeddings with logistic regression achieved a macro F1 of 0.981, near-perfect classification on formal financial language.
FinBERT AI Statistics 2026 on Training Data
The yiyanghkust FinBERT release was pretrained on 4.9 billion tokens of financial communication text. The corpus split is documented across three text types.
| Corpus | Token Count | Share |
|---|---|---|
| Corporate 10-K and 10-Q filings | 2.5 billion | 51% |
| Earnings call transcripts | 1.3 billion | 27% |
| Analyst reports | 1.1 billion | 22% |
| Total | 4.9 billion | 100% |
Source: Huang, Wang and Yang, Contemporary Accounting Research
FinBERT2, released by ValueSimplex at KDD 2025, scaled this approach to a 32 billion token finance-specific pretraining corpus. The new model targets dense retrieval and topic modelling alongside sentiment classification.
FinBERT AI Statistics 2026 on Fine-Tuned Variants
The base ProsusAI/finbert model has spawned 87 publicly listed fine-tuned variants and 7 adapter models on Hugging Face. The yiyanghkust release lists 18 finetunes covering ESG classification, forward-looking statement detection, and FOMC minutes analysis.
| Base Model | Fine-Tunes | Adapters | Quantizations |
|---|---|---|---|
| ProsusAI/finbert | 87 | 7 | 3 |
| yiyanghkust/finbert-tone | 18 | 1 | 0 |
Source: Hugging Face Model Hub
Common Specialised Variants
FinBERT-FOMC fine-tunes the base model on Federal Open Market Committee minutes from 2006 to 2023, achieving higher accuracy than the original on complex monetary policy sentences. FinBERT-ESG targets environmental, social and governance disclosures. ModernFinBERT, built on the ModernBERT backbone, reports up to 48% accuracy gains on multi-domain financial text including crypto and macroeconomic news.
FinBERT AI Statistics 2026 on Industry Adoption
Documented institutional users of FinBERT include Prosus, ING, Moody’s Analytics, and the London Stock Exchange Group. LSEG integrates the yiyanghkust/finbert-fls variant into developer tools for tagging forward-looking statements in filings.
Application areas span sentiment scoring of earnings calls, automated tagging of 10-K disclosures, ESG topic detection, and signal generation for systematic trading desks.
| Organisation | Use Case Category |
|---|---|
| Prosus AI | Original developer, internal research |
| Hugging Face | Model hosting and inference API |
| Moody’s Analytics | Risk and credit analysis |
| ING | Banking text analytics |
| LSEG (Refinitiv) | Forward-looking statement tagging |
Source: ProsusAI FinBERT documentation
FinBERT vs Other Models: 2026 Comparison
A peer-reviewed 2024 study published in Big Data and Cognitive Computing benchmarked FinBERT against GPT-4 and logistic regression on stock sentiment classification. Logistic regression posted 81.83% accuracy and a ROC AUC of 89.76%, while GPT-4 in zero-shot mode scored 54.19%.
FinBERT delivered moderate accuracy with a higher computational cost than logistic regression, but stronger contextual handling than the simpler model on technical financial vocabulary.
| Model | Accuracy | ROC AUC | Compute Cost |
|---|---|---|---|
| Logistic Regression | 81.83% | 89.76% | Low |
| FinBERT | Moderate | Moderate | High |
| GPT-4 (zero-shot) | 54.19% | Lower | Highest |
Source: Shobayo et al., Big Data and Cognitive Computing, MDPI
FinBERT AI Statistics 2026 on Model Size and Architecture
The standard FinBERT model has 110 million parameters, identical to BERT-base. The architecture uses 12 transformer layers, 768-dimensional hidden states, and a 30,522 token WordPiece vocabulary.
FinBERT outputs softmax probabilities over three labels: positive, negative, and neutral. The Xenova ONNX port reduces deployment overhead and runs in browsers through the Hugging Face Transformers.js library.
| Specification | Value |
|---|---|
| Parameters | 110 million |
| Transformer layers | 12 |
| Hidden size | 768 |
| Vocabulary size | 30,522 tokens |
| Output classes | 3 (positive, negative, neutral) |
Source: Hugging Face Model Card, ProsusAI/finbert
The 2024 paper from Yanxin Shen and Pulin Kirin Zhang found that combining FinBERT with prompt-engineered LLMs in zero-shot and few-shot setups raised classification accuracy on financial reports above either approach used alone. Research interest is shifting toward hybrid pipelines where FinBERT handles fast classification and LLMs handle longer-form reasoning, a pattern that has appeared across AI future trends through 2030.
FinBERT AI Statistics 2026 on Research Citations
The original 2019 Araci paper on arXiv has been cited extensively across financial NLP research. The 2022 Huang, Wang and Yang paper in Contemporary Accounting Research is the most-cited finance-domain BERT publication, with hundreds of follow-up studies.
A 2025 Bloomberg meta-analysis of 681 financial NLP papers from 2022 to 2025 identified FinBERT as one of the dominant baselines across the field, alongside more recent additions like FinGPT and BloombergGPT.
Research Output by Year
| Year | FinBERT-Related Papers (Approx.) |
|---|---|
| 2020 | 45 |
| 2021 | 110 |
| 2022 | 180 |
| 2023 | 240 |
| 2024 | 290 |
| 2025 | 340 |
Source: Bloomberg MetaGraph meta-analysis, arXiv 2509.09544
The growth of FinBERT-related work tracks the broader rise in generative AI research output and overlaps with AI applications in finance. Hybrid pipelines are also reshaping algorithmic trading, with FinBERT supplying sentiment features to downstream LSTM and transformer models.
FAQ
How many monthly downloads does FinBERT have on Hugging Face?
The ProsusAI/finbert model recorded 3,508,680 downloads in the last month. The yiyanghkust/finbert-tone variant added 947,386 downloads, putting combined monthly usage of the top two models above 4.4 million.
What accuracy does FinBERT achieve on financial sentiment tasks?
FinBERT reaches 97% accuracy on the full-agreement subset of the Financial PhraseBank dataset and 86% on the full dataset. Fine-tuned hybrid versions can reach a macro F1 of 0.981 on the same benchmark.
How large is the FinBERT training corpus?
The yiyanghkust FinBERT was trained on 4.9 billion tokens, split across 2.5B from 10-K and 10-Q filings, 1.3B from earnings calls, and 1.1B from analyst reports. The newer FinBERT2 uses 32 billion tokens.
How many fine-tuned FinBERT models exist?
Hugging Face lists 87 fine-tuned variants and 7 adapters built on ProsusAI/finbert. The yiyanghkust release has 18 finetunes covering ESG, FOMC minutes, forward-looking statements, and other specialised tasks.
Which companies use FinBERT in production?
Documented users include Prosus, Hugging Face, Moody’s Analytics, ING, and the London Stock Exchange Group. LSEG integrates a FinBERT variant into its developer tools for tagging forward-looking statements in regulatory filings.
Sources
