# Bloomberg's AI is Live. It Just Isn't BloombergGPT https://belitsoft.com/bloomberggpt ## **BloombergGPT Is in Production or Not?** In March 2023, Bloomberg published [a paper](https://arxiv.org/abs/2303.17564) about BloombergGPT, "a 50 billion parameter language model that is trained on a wide range of financial data". The training set was "a 363 billion token dataset based on Bloomberg's extensive data sources, perhaps the largest domain-specific dataset yet, augmented with 345 billion tokens from general purpose datasets". Bloomberg claimed that BloombergGPT "outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks". After the paper, Bloomberg gave no official updates about BloombergGPT. You could not find a GitHub repo or a [Hugging Face](https://belitsoft.com/ai-software-development-services/hugging-face) model card with a downloadable checkpoint, no public demo was available, and the price was not announced. However, only two months after Bloomberg's paper, a general model available through an API beat BloombergGPT on most of the tested financial tasks. Researchers from Queen's University and J.P. Morgan AI Research [tested ChatGPT against BloombergGPT](https://arxiv.org/abs/2305.05862) on financial tasks. According to their findings, ChatGPT performed "better on financial NLP tasks than BloombergGPT, which was specifically trained on financial corpora". On question answering, it was ahead of BloombergGPT by more than 15 percentage points. Despite this criticism, Bloomberg kept silent. Nine months later, Doug Levin, founder and first CEO of Black Duck Software and now an Executive Fellow at Harvard Business School, wrote [a review](https://douglevin.substack.com/p/bloomberggpt-a-new-era-in-real-time) in February 2024: "I recently had the opportunity to use BloombergGPT on the Bloomberg Professional service. The way I understand it, BloombergGPT was developed by Bloomberg's ML Product and Research group collaborating with the firm's AI Engineering and data teams." His experience was very positive: the AI was able to quickly generate high-quality financial reports, market summaries and trading ideas or analysis that incorporated real-time data from the platform. He wrote that "the potential for augmented intelligence in finance is clear", and that "BloombergGPT has disrupted the Bloomberg terminal with a new technology that is likely to replace the established one". Directly mentioned use cases from Doug Levin's article: 1. Financial report generation 2. Market summaries 3. Trading ideas or analysis 4. Financial data analysis 5. Market trend predictions 6. Sentiment analysis 7. Automated report generation 8. Language translation 9. Financial document text generation 10. Risk assessment 11. Real-time market updates 12. Support in client communication 13. Support in regulatory compliance 14. S-1 analysis and modeling 15. Search functionality via Bloomberg Search (SEAR) It seemed that, despite the skepticism of the researchers, Bloomberg had rolled out its model in its products and the clients were very happy. Bloomberg had a custom model, and this was now the only way to build a financial LLM that stands out. This opinion held in the community until 2026. However, in April 2026, Wayne Barlow, Global Head of Terminal Products at Bloomberg, told [A-Team Insight](https://a-teaminsight.com/blog/inside-bloombergs-askb-roadmap-the-data-moat-behind-the-ai/) that "BloombergGPT was a research model and is actually not used at all in any of our products". In fact, their products "are built on a combination of models". They have "some commercial models, some open-weight LLMs, some smaller, custom language models that we build ourselves. Because we have an agentic system, different layers may use different models, each solving specific needs. Some of it relates to latency and speed, in which case you may need a smaller model. In other cases, it's simply which model performs best for a specific task." Shawn Edwards, Bloomberg's CTO, confirmed this. He [told Fortune](https://fortune.com/2026/04/28/bloomberg-askb-ai-agents-lessons-from-bloomberg-cto-shawn-edwards-eye-on-ai/) that the system uses "a variety of AI models, including some built by Bloomberg itself and others from frontier AI model companies such as Anthropic", routing queries "to the cheapest model that can handle a given task". So Bloomberg trained its own large financial LLM (around 1.3 million GPU hours, a core team of about nine people), but this particular model never went into a product. They learned how to build one mega custom model, and then applied this knowledge to build small custom models for specific tasks, and only for the tasks where the ready-made LLM models did not have enough capabilities. Where a big commercial model did a task at a higher cost than a cheaper model, the task went to the cheaper model. In fact, they built the conversational AI interface ASKB, which uses different models depending on the task, and this matches the way financial AI products are built now. ### **BloombergGPT-like LLM: Train from Scratch or Use the APIs of Public Commercial Models?** In July 2023, four months after the paper, [David Rosenberg](https://www.linkedin.com/in/dr-david-rosenberg), who leads the Machine Learning Strategy team in the Office of the CTO at Bloomberg and is a co-author of BloombergGPT, told [The TWIML AI Podcast](https://www.youtube.com/watch?v=byCe7-c84d4) with Sam Charrington how the model was built. This is the only detailed description of the custom LLM development process from Bloomberg. Then Rosenberg explained why Bloomberg trained its own model this way. "Using an [API like OpenAI’s](https://belitsoft.com/openai-api-integration-services) is not suitable for us: we have data we don’t want to send out. So for internal and sensitive use, in-house models are preferable." — David Rosenberg, *The TWIML AI Podcast* It seems that in 2026 Bloomberg's position was clarified and adjusted, and in some cases they are no longer against using external LLM APIs and do not worry about security. Nevertheless, it looks like the most critical data still stays inside anyway, and it is processed by custom self-hosted models based on what the team learned from BloombergGPT. The results of BloombergGPT are outdated, but the cutom development process is not, and it is still useful for any company that plans its own financial LLM. ## **Backstage of Bloomberg's Custom Financial LLM Development** ### Decision Behind Creating a Custom Financial LLM BloombergGPT is an example of a project where a team inside an enterprise trained and built a custom large language model specializing in financial language. The enterprise made a strategic decision to invest money, time, and human resources into this machine learning effort. In 2026, as Barlow said, the model itself is not used in any product, and what they got from this investment is the experience of training a large model, the data pipeline and the team, which they now use for small custom models. "It was clearly a big investment. We decided it was worth to start building our own model. I'm not sure we knew exactly at that time what it would be used for." — — David Rosenberg, *The TWIML AI Podcast* ### Financial LLM use cases What the BloombergGPT development team actually spent the most time on was thinking about financial LLM use cases from a variety of angles: could the model help them solve problems they already had solutions for, but in a better way, or with less investment in training data? They explored several use cases, including the following ones: 1. Natural language to BQL. Bloomberg Query Language is used inside Bloomberg Terminal to pull structured financial data, and the idea was to build a kind of financial code assistant that translates human language into Bloomberg-specific queries. 2. Interaction with a large document. The user inputs a document and asks what information it contains. In fact, these particular two use cases became products much later. In April 2025, Bloomberg [launched Document Insights](https://www.prnewswire.com/news-releases/bloomberg-accelerates-financial-analysis-with-gen-ai-document-insights-302421875.html), where users "ask questions in everyday conversational language" about a company document, such as an earnings call transcript, and get "transparency links to highlight relevant excerpts in the original document". In February 2026, their financial AI product ASKB started to generate BQL. According to [The TRADE](https://www.thetradenews.com/bloomberg-embeds-agentic-ai-into-the-terminal/), "where queries involve data analysis, the system generates the associated Bloomberg Query Language code, allowing traders to extend outputs directly into Excel, BQuant Desktop or BQuant Enterprise for further modelling and workflow integration". ### LLM Hallucination problem "As for production use, we need to be very cautious. No one has really solved the hallucination problem yet. These language models can say wrong things, do strange things, so there needs to be a process around making them safe to use, either internally or, eventually, with clients." — David Rosenberg, *The TWIML AI Podcast* Moreover, if an LLM product seems reliable at first, the users who rely on it for everyday internal tasks may become less critical of its output in the future, and if the LLM eventually returns something wrong to the people who trust it, this can end in disaster. So the hallucination problem has to be solved at the system level. However, hallucinations were the second priority at the start of the custom financial LLM development. They started with internal use, paying less attention to safety or reputation and more to the development and testing of the functionality. The main question was: did it do the job? That was their focus at the time. Nevertheless, the question of solving hallucinations was still there. The simple solution seemed to be that the user just checks the output for hallucinations. But after this came another strategic question: how much of the responsibility for checking should be moved to the user, and will it turn out in the end that the users spend as much time or more on checking, and there is no efficiency gain that is expected from using an LLM? At that time, they had no answer to this question, so they kept things internal. However, by 2026 it became known how Bloomberg solved the hallucination problem and built "a useful, trustworthy system" (Shawn Edwards, Bloomberg's CTO). 1. Evaluations at the product development level. As Edwards [told Fortune,](https://fortune.com/2026/04/28/bloomberg-askb-ai-agents-lessons-from-bloomberg-cto-shawn-edwards-eye-on-ai/) using AI models as evaluators "can work for easy cases", and "for everything else, human assessors are required". During development, Bloomberg pulled bond covenant experts, equity analysts, market structure specialists and journalists off their day jobs both to write benchmarks for sub-agents and to help evaluate entire workflows". 2. Checking at the user level, which is still cheaper fro business users than collecting the information yourself. According to [Bloomberg's press release](https://www.bloomberg.com/professional/insights/press-announcement/bloomberg-unveils-askb-roadmap-for-clients-to-augment-their-investment-process-with-agentic-ai/), every ASKB answer comes with "transparent attribution to original sources and the underlying Bloomberg Query Language code for any data analysis". ### Training Dataset for Bloomberg AI The training dataset was a mix of standard general-purpose data and Bloomberg's proprietary financial data. About half of the dataset (51.27%) came from Bloomberg's curated collection called FinPile, built over many years starting in 2007. Most of FinPile (298 of 363 billion tokens) was financial content scraped from the web. The rest was news articles, filings, press releases, and Bloomberg's own content. There was a problem with the tokenization of numerical data, and they solved it. In 2026, this problem is solved at the tokenizer level in most open-weight models, so a company that fine-tunes a ready-made model does not need to do this work again. The BloombergGPT model did not go into a product, but the data did. In 2026, ASKB draws on "hundreds of millions of company documents like earnings call transcripts, presentations and filings", and Shawn Edwards told Fortune that the data is the hard part: "You have to buy all those sources, do all the validation work, build benchmarks, and tokens aren't cheap." ### Custom BloombergGPT Training Process At that time, Meta had just released the OPT model together with its training logs, and Hugging Face had published the detailed logs of BLOOM. The team used these logs as a roadmap. "We copied the BLOOM model architecture fairly closely, with some small tweaks. " — David Rosenberg, The TWIML AI Podcast The training took three attempts: 1. v0, curriculum learning. FinPile data was sorted chronologically by month, because the newer data was more similar to the future documents, and the validation set was the month after the training set. After about a week, the model was stuck on both training and validation loss, and they stopped this run, because they "deemed this to be too risky to catch any other potential problems with the training that might require early intervention". They restarted with fully shuffled data. 2. v1, gradient norm growth. After about 4.5 days (12k steps) the gradient norm started to grow, and the validation loss jumped. They made four attempts to fix it (a rollback to an earlier checkpoint, reshuffled data, a lower learning rate, gradient clipping, FP32 in the LM head), and none of them worked. Then they found that out of 70 layers, the layer norm weights of the first layer shrank until steps 11-12k and then started to grow. During this investigation, they also found a bug: weight decay was applied to the layer norm weights that should stay around one. 3. v2, the final run. They started from scratch with conservative hyperparameters (a lower learning rate, gradient clipping 0.3, a longer warm-up, an extra layer norm at the embedding layer, no weight decay on layer norm weights, fully shuffled data) and additional monitoring of the first layer. The training was "very smooth and uneventful" for about 42 days. Around day 48, the validation loss stopped improving for a week, they tried six more runs with a lower learning rate, a rollback and reshuffled data, and stopped training after 53 days at about 80% of the dataset, because "the loss on our held-out development set was no longer improving". ### Resources Used to Train BloombergGPT The core team was about nine people (four on implementation, three on ML and data, one on optimization and compute, and the rest on evaluation, literature review and support). The compute was 512 40GB A100 GPUs on Amazon SageMaker (64 p4d.24xlarge instances with 8 GPUs each) with the SageMaker Model Parallelism library. Bloomberg pre-purchased around 1.3 million GPU hours at a negotiated rate. They took this compute budget and calculated from it, by the Chinchilla scaling laws, how big a model (50 billion parameters) and how much training data (569 billion tokens) fit into 1.3 million GPU hours. So the whole project was about $1.3 million at the low end of cloud pricing ($1 per hour for reserved capacity). ### Custom Financil LLM Validation and Performance Evaluation During training, they used the month after the training data for validation. Later, they added a second validation set from the last 105 million tokens of the shuffled training data. They also monitored downstream tasks like MMLU and BBH (Big Bench Hard). After training, they did a full evaluation and compared BloombergGPT to other models. On financial tasks, it outperformed the open models of that time. They also had internal benchmarks: 1. Sentiment analysis on financial news and social media 2. Named entity disambiguation (linking "Apple" to its stock ticker, etc.) 3. Natural language to BQL, which is like SQL for Bloomberg Terminal Even though the model wasn't trained on BQL directly, it performed well in few-shot scenarios. They also experimented with headline generation and other generative tasks. In 2026, the evaluation is the main part of the work. The benchmarks are now written by domain experts, and not only by the ML team. ### What’s Next for the BloombergGPT Team? Already in 2023, with a working large model, they understood that they needed to move to small-scale experiments and test tokenization strategies, data mixtures and architecture choices in a more disciplined way. "We're more interested now in smaller models. They're easier to use: you can run inference on a single GPU. Our 50B model requires multi-GPU infrastructure. Inspired by the LLaMA paper, we're exploring what we can achieve with smaller models, longer training, and careful design. We want both small and large models for practical use." — David Rosenberg, The TWIML AI Podcast They managed to move in this direction and stick with it. In 2026, Bloomberg's products run on "smaller, custom language models that we build ourselves" together with commercial and open-weight models, as Barlow said. The 50B model never went into a product. They also continued the small-scale experiments. On July 2, 2026, Bloomberg's AI researchers [presented 8 papers at ACL 2026](https://www.bloomberg.com/company/stories/bloombergs-ai-researchers-publish-8-papers-at-acl-2026-spanning-agentic-systems-responsible-ai-and-financial-nlp/), including "a new training recipe that enables smaller models to match or outperform larger specialized guardrails on unseen policies" and "a novel approach to complex text-to-SQL". ## **Your own Financial LLM: A Mirage?** Financial LLMs are not a fantasy. There are already startups in this space that have gone to production, raised funding, and landed their first clients. The most well-known are those backed by YCombinator, for example, [Truewind](https://belitsoft.com/truewind-ai-case-study-finance-accounting) ($17 million raised, Series A in January 2025, over 100 customers in 2026). However, Truewind did not train its own model and uses models from OpenAI, as well as JPMorgan's LLM Suite, launched in 2023, which uses models from OpenAI and Anthropic. Goldman Sachs built its GS AI Assistant so that it can interact with different LLMs from OpenAI, Google and Anthropic, as well as open source models. The model vendors came to finance themselves. On May 5, 2026, Anthropic [released](https://www.anthropic.com/news/finance-agents) "ten ready-to-run agent templates for the most time-consuming work in financial services". Today, Claude connects to dozens of market data, research platforms, and financial companies' internal systems, including FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, Chronograph, LSEG, and Daloopa. The financial models became small. [Fin-R1](https://arxiv.org/abs/2503.16252) (latest version March 2026) has "a compact size of 7 billion parameters" and, despite it, achieves competitive performance across financial benchmarks. The mirage is one big custom financial LLM. Bloomberg trained one, and in 2026 this model is not used in any of its products. LLMs also have core limitations by design, and you can't expect a financial LLM to do what it simply wasn't built to do. So, one big custom financial LLM is a mirage. Small custom models for specific tasks, together with pre-built commercial models, are not. ## **How Belitsoft Can Help** Building a financial LLM is not a technical challenge. It's a product challenge. Belitsoft helps financial firms build AI products on a combination of models, as Bloomberg did with ASKB, where each task goes to the model that handles it at the lowest cost. This means: - Training small custom models for specific tasks, where ready-made models do not have enough capabilities - Fine-tuning existing open-weight models with in-house financial content - Connecting commercial models (OpenAI, Anthropic, Google) for the tasks where they are cheaper or better - Routing tasks between these models depending on latency and cost - Evaluations at the product development level by your domain experts - Checking at the user level: every answer comes with its sources and the underlying query - A conversational AI interface for your users or an integration into your dashboards and query layers You get your own financial AI product: an internal assistant, a decision tool or a document engine. The most critical data stays inside and is processed by custom self-hosted models. The other tasks go to pre-built commercial models through their APIs. ## Frequently Asked Questions ![](https://belitsoft.com/uploads/images/design_landing/point.svg) Is BloombergGPT open source?![](https://belitsoft.com/belitsoft_refresh/dist/images/php-landing/plus.svg) BloombergGPT is not open-source. After the paper in March 2023, Bloomberg gave no official updates about BloombergGPT. You could not find a GitHub repo or a Hugging Face model card with a downloadable checkpoint, no public demo was available, and the price was not announced. Bloomberg has [an organization on Hugging Face](https://huggingface.co/bloomberg), but there are only two keyphrase extraction models from 2023 and one dataset from 2022. The architecture was based on BLOOM, but the model itself is proprietary. In April 2026, Wayne Barlow, Global Head of Terminal Products at Bloomberg, said that "BloombergGPT was a research model and is actually not used at all in any of our products". So you cannot build on BloombergGPT. In fact, you do not need your own large model, because Bloomberg itself runs its products on a combination of commercial models, open-weight models and small custom models for specific tasks. ![](https://belitsoft.com/uploads/images/design_landing/point.svg) BloombergGPT Cost![](https://belitsoft.com/belitsoft_refresh/dist/images/php-landing/plus.svg) Bloomberg did not publish the cost of BloombergGPT. However, the numbers for the calculation are public. **Cost of Training** In 2022, Bloomberg pre-purchased around 1.3 million GPU hours at a negotiated rate, and the model was trained on 512 A100 40GB GPUs on Amazon SageMaker. At the low end of cloud pricing ($1 per hour for reserved capacity), the training cost about $1.3 million. **Cost of Development** In late 2020, Bloomberg decided to build the model, and in March 2023 the paper came out. So the core team of about nine people worked on the model for more than two years. It seems that with senior-level salaries in the US the development cost between $4.5 million and $7 million, and the full cost of BloombergGPT with training is between $6 million and $8.5 million. Bloomberg did not confirm these numbers. **Cost of Using** In April 2026, Barlow said that the model is not used in any of Bloomberg's products, so you cannot use BloombergGPT at any price. The AI that you can use is ASKB, and it is a part of the Terminal. The Terminal starts at $31,980 per user per year in 2026 and accounts for over 85% of Bloomberg's total annual revenue. ASKB has no separate price, and it seems that Bloomberg does not sell it, but uses it to keep the Terminal subscribers.