Rapid advancements of open-source models are prompting financial institutions to pull back on frontier-model usage and reevaluate their AI strategies.  

Open-weight models are becoming “increasingly capable and are putting downward pressure on AI pricing, which is good news for all LLM users, including financial institutions,” Chris Griffin, co-founder of digital banking platform Narmi, told FinAi News.  

Open-source models come with complete source codes and data sets; open-weight models provide the downloadable parameters without the entire underlying code or data.  

“While many financial services use cases require the most advanced models available, there are also high-volume, more straightforward applications, such as transaction enrichment, where a less expensive open-weight model could be more than capable,” he said.  

“As these models continue to improve, they could give financial institutions more cost-effective options for deploying AI at scale.”

— Chris Griffin, co-founder, Narmi

Thread Bank, for example, has ramped up experimentation with open-source models over the past few months, using LLMs offline after downloading them to Apple Mac Studio, Chief Digital Officer and Chief Risk Officer Marty Miracle told FinAi News.  

The bank has found some models “more useful, cheaper or just as good” compared with frontier models, Miracle said. In fact, the $1 billion bank recently built a marketing compliance solution through an open-source model, he added.  

Open models can cost 10 to 50 times less than frontier models, Aser Blanco, global head of banking and financial services at Nvidia, recently told FinAi News. 

Open-source models are publicly available for download and offer high customizability, whereas frontier models are proprietary, more advanced systems that can only be accessed via API, according to the Cloud Security Alliance. 

Common open-source models include:

• Meta’s LLama series;
• Google’s Gemma series;
• DeepSeek’s R1 series; and
• Mistral AI’s Ministral 3. 

Striking balance  

Financial institutions constantly must reevaluate their mix of open and frontier models as the technologies evolve at a rapid pace, Andrew Moore, chief executive of risk management fintech Lovelace AI and former head of Google Cloud AI, told FinAi News.  

FIs “do still need frontier models, but they need them rarely,” he said.  

For example, the most powerful Anthropic Claude models are not needed nearly as much for coding today compared with six to 12 months ago, he said.  

Financial institutions should increasingly lean on model routing techniques so that simpler tasks are directed to open models, while frontier models handle more complex tasks, Sahil Sagar, chief information and AI officer at real estate lender Greystone, told FinAi News. 

“You do not need the most powerful model from Anthropic to rewrite your email,” he said. “That should be a much cheaper model.” 

Global competition  

Chinese LLM provider DeepSeek released V4 Pro and V4 Flash over the past several weeks, its most advanced open models, according to the company. 

Other Chinese open-source LLM providers include Moonshot AI and Z.ai, with France-based Mistral AI emerging as another global competitor.  

Many leading open-weight models are “developed outside the U.S., which can create regulatory considerations for U.S. financial institutions,” Narmi’s Griffin said. 

“However, the emergence of new U.S.-based open-weight models from companies like [Thinking Machine Labs] and Meta is an encouraging development, giving financial institutions more domestic options as they evaluate how to deploy AI,” he said. 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas. This inaugural event will include speakers from Fifth Third and Capital One as well as a fireside chat with Piermont Bank founder and Chief Executive Wendy Cai-Lee.