BloombergGPT was trained on 709 billion tokens — 363 billion of them from Bloomberg’s proprietary financial archive — making it one of the largest domain-specific language models ever built. The 50.6 billion parameter model cost between $2.67 million and $10 million to train and recorded 62.5% average accuracy on public financial benchmarks, outperforming similarly sized open models by 8 to 10 percentage points. This article covers BloombergGPT’s architecture, benchmark performance, training data composition, enterprise integration through Bloomberg’s ASKB platform, and where the model sits in the broader financial AI market heading into 2026.
BloombergGPT Statistics in 2026 — TL;DR
BloombergGPT contains 50.6 billion parameters across 70 transformer layers, trained on 709 billion tokens with 51.27% from Bloomberg’s proprietary FinPile dataset.
Training consumed 1.3 million GPU hours on 512 NVIDIA A100 GPUs over 53 days.
The model recorded 62.5% average accuracy on public financial datasets versus 51.9%–54.4% for competing models of similar size.
Bloomberg has over 325,000 Terminal subscribers worldwide as of 2026, with each subscription costing approximately $31,980 per year.
The generative AI in financial services market reached $1.95 billion in 2025 and is projected to grow to $7.24 billion by 2030 at a 30.7% CAGR.
BloombergGPT is a decoder-only causal language model purpose-built for financial NLP tasks. Bloomberg released the research paper in March 2023, and since then, the model has been integrated into workplace AI tools within the Bloomberg Terminal ecosystem. In February 2026, Bloomberg launched ASKB, an agentic AI interface that lets Terminal users interact with financial data using natural language — a direct evolution of the capabilities BloombergGPT introduced.
How Large Is BloombergGPT’s Architecture?
BloombergGPT runs on a decoder-only transformer with 50.6 billion parameters. The model uses 70 layers, 40 attention heads, and a hidden dimension of 7,680. Its vocabulary contains 131,072 tokens — more than double the standard 50,000-token BPE vocabulary used in most general-purpose models. This larger vocabulary was designed to better capture financial terminology, ticker symbols, and numerical structures common in financial documents.
The architecture uses ALiBi positional encoding, which allows the model to generalize to longer sequences than it was trained on. Bloomberg’s ML Product and Research team selected these parameters based on Chinchilla scaling laws, balancing model size against available compute within a fixed budget of 1.3 million GPU hours.
| Specification | Value |
|---|---|
| Total Parameters | 50.6 Billion |
| Transformer Layers | 70 |
| Attention Heads | 40 |
| Hidden Dimension | 7,680 |
| Vocabulary Size | 131,072 Tokens |
| Positional Encoding | ALiBi |
| Architecture Type | Decoder-Only Causal LM |
Source: Bloomberg Research Paper (arXiv:2303.17564)
What Data Was BloombergGPT Trained On?
The training corpus totaled 709 billion tokens. Bloomberg constructed FinPile, a proprietary dataset of 363 billion tokens drawn from financial documents spanning 2007 to 2022. This included news articles, SEC filings, press releases, web-scraped financial content, and social media posts from Bloomberg’s archive. The remaining 345 billion tokens came from public datasets — The Pile, C4, and Wikipedia.
FinPile accounted for 51.27% of the total training data, with public sources making up the other 48.73%. Bloomberg sorted FinPile chronologically and randomly mixed it with the public data during training. The team used 512 NVIDIA A100 40GB GPUs and completed training in 139,200 steps over 53 days.
BloombergGPT Training Data Composition
| Data Source | Tokens (Billions) | Share of Corpus |
|---|---|---|
| FinPile (Proprietary) | 363 | 51.27% |
| The Pile | ~184 | ~26% |
| C4 | ~138 | ~19.5% |
| Wikipedia | ~23 | ~3.2% |
Source: Bloomberg Research Paper (arXiv:2303.17564)
BloombergGPT Benchmark Performance
On public financial datasets — including FPB, FiQA SA, Headline, Fin-NER, and ConvFinQA — BloombergGPT recorded an average accuracy/F1 score of 62.5%. Competing models in the same size range (GPT-NeoX, OPT-66B, BLOOM-176B) scored between 51.9% and 54.4% on the same tasks. On named entity recognition and disambiguation tasks, BloombergGPT showed a 25 to 30 point gap over GPT-NeoX and OPT-66B.
On FiQA sentiment analysis specifically, BloombergGPT hit 75.07% F1 compared to GPT-NeoX’s 50.59% — a 25-point lead. The model also performed at or above peer level on general-purpose benchmarks like BIG-Bench Hard, MMLU, and reading comprehension tasks, which means the financial specialization did not come at the expense of broader generative AI capabilities.
BloombergGPT vs. Competing Models on Financial Tasks
| Benchmark | BloombergGPT | GPT-NeoX | OPT-66B |
|---|---|---|---|
| FiQA Sentiment (F1) | 75.07% | 50.59% | 52.10% |
| Avg. Financial Tasks | 62.5% | 51.9% | 54.4% |
| NER/NED (Relative) | +25–30 pts lead | Baseline | Baseline |
Source: Bloomberg Research Paper; Beancount.io Analysis (May 2026)
How Does BloombergGPT Compare to GPT-4?
GPT-4 outperformed BloombergGPT on most financial tasks despite not having domain-specific pretraining. On FinQA, GPT-4 scored 68.79% compared to BloombergGPT’s lower results on the same benchmark. GPT-4 contains an estimated 1 trillion parameters and was trained on roughly 100 times the compute used for BloombergGPT.
This result raised questions about whether domain-specific pretraining remains valuable when general-purpose models keep scaling up. For ChatGPT and similar platforms with far larger parameter counts, raw model scale can close the gap that domain specialization creates. Bloomberg appears to have acknowledged this by shifting its AI strategy toward ASKB, which uses retrieval-augmented generation rather than relying solely on BloombergGPT’s internal knowledge.
| Attribute | BloombergGPT | GPT-4 |
|---|---|---|
| Parameters | 50.6 Billion | ~1 Trillion (est.) |
| Training Tokens | 709 Billion | ~13 Trillion (est.) |
| FinQA Accuracy | Lower | 68.79% |
| Domain Data | 363B proprietary tokens | None (general web) |
| Training Cost | $2.67M–$10M | $100M+ (est.) |
Source: Bloomberg Research; XYZEO Review (Feb 2026)
BloombergGPT Training Cost and Infrastructure
Training BloombergGPT cost between $2.67 million and $10 million, depending on the cost model applied to GPU hours. The team used 512 NVIDIA A100 GPUs with 40GB of memory each, parallelized across multiple machines. The full training run took 53 days and consumed 1.3 million GPU hours.
By comparison, GPT-4’s training cost exceeded $100 million according to industry estimates. BloombergGPT’s comparatively modest budget reflects the trade-off Bloomberg made: a smaller model with specialized data rather than a massive general-purpose system. The training process included activation checkpointing to manage memory and incorporated lessons from earlier open-source training efforts like OPT and BLOOM.
Training Cost Comparison: Domain-Specific vs. General LLMs
Bloomberg Terminal and Enterprise Integration
Bloomberg has over 325,000 Terminal subscribers worldwide as of 2026, generating roughly $11 billion in annual revenue. A single Terminal subscription costs $31,980 per year, with multi-seat pricing dropping to $28,320 per terminal. The subscriber base spans investment banks, asset managers, hedge funds, pension funds, and government agencies across more than 170 countries.
In February 2026, Bloomberg launched ASKB — a conversational AI interface that lets users query structured data and unstructured documents using natural language. ASKB generates Bloomberg Query Language (BQL) from plain-English requests, consolidating what previously required navigating multiple Terminal screens. In April 2026, Bloomberg released a roadmap expanding ASKB into firm-wide intelligence, integrating AI with existing data privacy frameworks and proprietary research from providers like Third Bridge.
| Bloomberg Terminal Metric | 2026 Value |
|---|---|
| Global Subscribers | 325,000+ |
| Annual Revenue | ~$11 Billion |
| Single Seat Price | $31,980/year |
| Multi-Seat Price | $28,320/year |
| Countries Served | 170+ |
| AI Interface | ASKB (launched Feb 2026) |
Source: Bloomberg Professional Services; Comstock Interactive Data (April 2026)
Generative AI in Financial Services Market Size
The generative AI in financial services market was valued at $1.95 billion in 2025. Projections put it at $2.48 billion in 2026, growing at a 31.1% CAGR. By 2030, the market is expected to reach $7.24 billion. North America held 42% of the market revenue in 2024, and risk management applications accounted for the largest segment at 27.9%.
The broader LLM market reached $8.33 billion in 2025 and is projected to hit $10.97 billion in 2026. Sub-100 billion parameter models — the category BloombergGPT falls into — captured 69.2% of market share in 2025. The enterprise LLM market specifically was valued at $4.84 billion in 2025, with banking and financial services accounting for a sizable portion of that spending. These numbers provide context for where BloombergGPT sits within the growing AI-driven financial sector.
Generative AI in Financial Services Market Growth
| Year | Market Size (USD Billion) |
|---|---|
| 2024 | $1.67 |
| 2025 | $1.95 |
| 2026 (Projected) | $2.48 |
| 2028 (Projected) | $4.20 |
| 2030 (Projected) | $7.24 |
Source: Research and Markets; Precedence Research (2025–2026)
BloombergGPT Use Cases in Finance
Bloomberg designed BloombergGPT to handle a specific set of financial NLP tasks. The model performs sentiment analysis on earnings calls and news, classifies financial headlines, extracts named entities like company names and stock tickers, and answers questions about financial documents. One of its standout capabilities is translating natural language into BQL — Bloomberg’s proprietary query language for data access.
Within the Terminal, these capabilities feed into workflows that previously required manual multi-step research. An analyst who used to cross-reference SEC filings, news feeds, and market data across multiple screens can now describe their research objective in plain language. The ASKB interface processes that request through a network of AI agents that pull from Bloomberg’s data universe. This shift mirrors broader trends in enterprise technology adoption where conversational interfaces are replacing menu-driven navigation.
BloombergGPT Model Performance by Task Category
BloombergGPT’s strongest results came on financial sentiment and named entity recognition. On FiQA sentiment analysis, it achieved 75.07% F1. On the Headline dataset, it outperformed all open models under 176 billion parameters. For ConvFinQA (conversational financial question answering), the model showed solid gains but fell short of GPT-4-class performance.
On general NLP benchmarks, BloombergGPT performed on par with or slightly better than peer models. It matched BLOOM-176B on BIG-Bench Hard despite having roughly 3.5 times fewer parameters. Reading comprehension and linguistic tasks showed no degradation from the financial-domain training — a result that validated Bloomberg’s mixed-dataset approach.
BloombergGPT Accuracy Across Task Categories
Enterprise LLM Market and BloombergGPT’s Position
The global enterprise LLM market was valued at $4.84 billion in 2025, projected to reach $5.91 billion in 2026 and $48.25 billion by 2034 at a 30% CAGR. North America accounted for 42.7% of the market in 2025. The banking, financial services, and insurance (BFSI) sector is one of the largest verticals, with 71% of organizations now using generative AI in at least one business function according to McKinsey’s 2025 survey.
BloombergGPT occupies a niche within this market as one of the few domain-specific LLMs built by a financial data company using proprietary data. Its closest analogy would be if Reuters or S&P Global built their own language model trained on decades of proprietary financial records. The model’s long-term position depends on whether domain-specific pretraining continues to add value as general-purpose models grow larger and more capable.
| Market Segment | 2025 Value | 2026 Projected | CAGR |
|---|---|---|---|
| Global LLM Market | $8.33B | $10.97B | 31.8% |
| Enterprise LLM Market | $4.84B | $5.91B | 30.0% |
| Gen AI in Financial Services | $1.95B | $2.48B | 31.1% |
| Gen AI in Banking & Finance | $1.75B | $2.36B | 34.8% |
Source: Fortune Business Insights; Research and Markets; The Business Research Company (2025–2026)
FAQ
How many parameters does BloombergGPT have?
BloombergGPT contains 50.6 billion parameters distributed across 70 transformer layers with 40 attention heads and a hidden dimension of 7,680.
How much did it cost to train BloombergGPT?
Training cost between $2.67 million and $10 million, using 512 NVIDIA A100 GPUs over 53 days and consuming 1.3 million GPU hours.
Does BloombergGPT outperform GPT-4 on financial tasks?
No. GPT-4 outperforms BloombergGPT on most financial benchmarks despite lacking domain-specific pretraining, due to its much larger parameter count and training compute.
How many Bloomberg Terminal subscribers are there in 2026?
Bloomberg has over 325,000 Terminal subscribers worldwide as of 2026, generating approximately $11 billion in annual revenue at $31,980 per seat.
What is ASKB and how does it relate to BloombergGPT?
ASKB is Bloomberg’s agentic AI interface launched in February 2026. It builds on BloombergGPT’s NLP capabilities, letting users query Terminal data with natural language.
Sources:
https://arxiv.org/abs/2303.17564
https://www.bloomberg.com/company/press/bloomberggpt-50-billion-parameter-llm-tuned-finance/
https://www.bloomberg.com/professional/insights/press-announcement/bloomberg-unveils-askb-roadmap-for-clients-to-augment-their-investment-process-with-agentic-ai/
https://www.researchandmarkets.com/reports/6226395/generative-ai-in-financial-services-market-report
