Curiosity · AI Model
BloombergGPT
BloombergGPT is a 50-billion-parameter decoder-only LLM trained by Bloomberg on roughly 700 billion tokens — half public web text, half Bloomberg's own financial corpus (FinPile). Released with a 2023 paper, it became a reference example for finance-specialised domain models and strongly outperforms generalist LLMs on finance-NLP benchmarks.
Model specs
- Vendor
- Bloomberg
- Family
- BloombergGPT
- Released
- 2023-03
- Context window
- 2,048 tokens
- Modalities
- text
Strengths
- Trained on Bloomberg's unique financial corpus (FinPile)
- Beats generalist LLMs on finance NLP at similar parameter count
- Preserves competitive performance on general benchmarks
- Published paper and data mixture — a widely cited reference
Limitations
- Closed model — not available outside Bloomberg products
- 2048-token context is short versus modern generalist LLMs
- 2023-era knowledge and architecture — modern FinMA / FinGPT / GPT-5 in finance often surpass it on open benchmarks
- Terminal-only access model — cannot be fine-tuned externally
Use cases
- Financial news summarisation and classification
- Ticker and entity recognition in filings
- Sentiment over earnings calls and research notes
- Reference architecture for domain-specialised LLMs
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| FinPile internal eval | SOTA at release | 2023-03 |
| FPB sentiment F1 | ≈85 | 2023 |
Frequently asked questions
What is BloombergGPT?
BloombergGPT is a 50-billion-parameter large language model built by Bloomberg in 2023. It was trained on roughly 700 billion tokens — half public web text and half Bloomberg's proprietary financial corpus (FinPile).
Can external developers access BloombergGPT?
No — BloombergGPT is an internal model used inside Bloomberg products (Terminal features, research tooling). The weights are not published and there is no public API. External teams that need a finance-specialised LLM usually fine-tune open models instead.
Why is BloombergGPT still notable in 2026?
It is one of the first well-documented finance-domain LLMs from a major data vendor, and its paper sets out a template — FinPile data mixture, domain+general blending — that open-source projects like FinGPT and FinMA follow.
What are modern open alternatives?
Open or API-accessible finance LLMs include FinGPT, FinMA, Cohere finance embeddings, voyage-finance, and generalist frontier models like GPT-5 and Claude Opus 4.7 with finance fine-tunes or retrieval over financial corpora.
Sources
- BloombergGPT paper (arXiv) — accessed 2026-04-20
- Bloomberg — BloombergGPT press release — accessed 2026-04-20