Announcing the LLM Open Finance models

AGEFI and Dragon LLM are launching the LLM Open Finance Initiative: a new open-source effort to bring finance-ready language models to everyone. The first release of the LLM Open Finance suite includes two 8B-parameter models specialized for financial language, with strong support for English and French. Built on a diverse training dataset combining curated financial corpora and instruction data, these models are designed for financial reporting analysis, risk assessment, regulatory compliance, financial translation, financial sentiment analysis and retrieval-augmented applications. This work was developed with AGEFI as part of a France 2030 initiative supported by Bpifrance.

Highlights

• Consistent gains on financial tasks versus base models, with significant improvements on financial French acronyms and domain terminology understanding.
• One of the best open source models for financial translation across multiple language and financial document types.
• RAG-ready: the models have been trained on large amounts of RAG data.

More on the released models

We release two 8B parameters models based on Llama 3.1 and Qwen 3 , Both finetuned models preserve the features of their base model, meaning that, for instance, dynamically enabling (or disabling) the reasoning mode with the /think / /no_think tokens is still supported on the Qwen 3 based model. The models also preserve their multilingual capabilities, even though our additional finetuning focus mostly on English, French and German languages.

We built a balanced high quality dataset enabling high performance and accuracy on finance related tasks while preserving general domain capabilities. Our dataset include 54% of financial data, 20% of translation data, 16% of general domain data, 8% of RAG data, and 2% of math, reasoning and coding data, whose purpose is to maintain the initial model’s capabilities in these area.

We evaluated our models on various datasets and we are proud to show that :

1. Our models outperform general domain models on financial tasks and financial translation, hence validating our data curation pipeline
2. Our models also outperform high performing finance models, such as Salesforce/Llama-Fin-8b
3. Our models maintain strong general domain knowledge, making them an ideal and risk-controlled drop-in replacement of similarly sized models

We are releasing LLM Open Finance to the open-source community to democratize access to advanced financial AI models. This initiative aims to foster innovation by empowering users to explore new applications, reproduce research results, and fine-tune the models for specific use cases, ultimately driving progress in financial technology and economic research.

If you want to learn more about this project, do not hesitate to read our paper (https://arxiv.org/abs/2511.08621) or contact us directly.

Beyond small models: the LLM Pro Finance Suite

For teams needing more capacity and advanced reasoning, our more advanced models named LLM Pro Finance are available via commercial licensing:

• Gemma Pro Finance 12B: best for financial translation, batch/high-frequency processes, classification
• Qwen Pro Finance R 32B: best for ⇒ financial mathematics, code generation / agentic systems, structured output
• Llama Pro Finance 70B: best for ⇒ Conversational chat, Retrieval-Augmented Generation (RAG), content generation

Get started

• Hugging Face model cards https://huggingface.co/collections/DragonLLM/llm-open-finance
• GitHub: Cookbooks and evaluation scripts https://github.com/Dragon-LLM/llm-open-finance-cookbook
• Playground: try the LLM Pro Finance models in your browser https://demo.llmprofinance.com/