BloombergGPT

BloombergGPT is a 50-billion-parameter language model built specifically for finance, described in a research paper released on March 30, 2023 (arXiv:2303.17564). It was trained on a 363-billion-token dataset drawn from Bloomberg's proprietary financial data sources — described as perhaps the largest domain-specific dataset assembled at the time — augmented with 345 billion tokens of general-purpose data. Bloomberg reported that the mixed training gave it best-in-class results on financial NLP benchmarks (sentiment analysis, named entity recognition, news classification, question answering) without sacrificing general LLM benchmark performance.

Limitations: BloombergGPT was never released publicly — there is no GitHub repo, Hugging Face checkpoint, or public demo; it exists as a research artifact used inside Bloomberg's own products and the Terminal. A follow-up study by Queen's University and J.P. Morgan AI Research found ChatGPT beating BloombergGPT on most tested financial tasks, including question answering by over 15 percentage points, which undercut the paper's headline claims. Treat it as a landmark in domain-specific LLM research rather than something you can download or deploy.