Models and platforms · computed from bank profiles; nothing on this page is written by a model. All ranked answers: /banks/leaders
On the public record as of Oct 28, 2026, 14 of the 141 large banks profiled have a dated profile entry naming an open-weight model (Llama, Mistral) or open-source models in their AI stack. Ranked by matching entries, the top 5 are BNP Paribas (13 entries), Capital One (7 entries), HSBC (5 entries), BofA (4 entries) and Flagstar (4 entries). BNP Paribas's strongest evidence: HelloïZ virtual assistant at Hello bank! — The latest version, using generative AI (Mistral AI models) to understand and route requests with access to human… (use case, in production, May 26, 2026).
141 banks profiled (120 US, 21 non-US) · 14 with matching evidence · ranked answer, not a comparison
Which of the large banks use open-weight or open-source AI models?
| # | Bank | Matching entries | Strongest entry | Status | Date | On this site | Sources |
|---|---|---|---|---|---|---|---|
| 1 | BNP Paribas | 13 | HelloïZ virtual assistant at Hello bank! — The latest version, using generative AI (Mistral AI models) to understand and route requests with access to human… (use case, in production, May 26, 2026) | In production | May 26, 2026 | BNP Paribas profile → | BNP Paribas ↗ · BNP Paribas ↗ |
| 2 | Capital One | 7 | Proprietary LLMs customised from open-source models (use case, in production, Sep 24, 2026) | In production | Sep 24, 2026 | Capital One · technology stack → | Capital One ↗ · VentureBeat (sponsored by Capital One) ↗ |
| 3 | HSBC | 5 | Mistral self-hosted models (use case, rolling out, Dec 1, 2025) | Rolling out | Dec 1, 2025 | HSBC profile → | HSBC Holdings plc ↗ · Finextra ↗ |
| 4 | BofA | 4 | Orchestra model-routing layer — Sends simple classification tasks to approved open-weight models on the bank's own GPUs and harder reasoning to… (use case, in production, Oct 1, 2026) | In production | Oct 1, 2026 | BofA · technology stack → | Fortune ↗ · Fortune ↗ |
| 5 | Flagstar | 4 | StarIQ multi-model orchestration — Claude, Llama, Mistral and Titan behind one governed platform with consumption-based deployment (use case, in production, Jun 15, 2026) | In production | Jun 15, 2026 | Flagstar · technology stack → | Flagstar Bank ↗ · Banking Dive ↗ |
| 6 | HSBC USA | 4 | Self-hosted Mistral models (use case, announced, Dec 1, 2025) | Announced | Dec 1, 2025 | HSBC USA profile → | HSBC ↗ · Financial Times ↗ |
| 7 | PNC | 3 | Open-weight models in PNC data centres (use case, in production, Sep 14, 2026) | In production | Sep 14, 2026 | PNC · technology stack → | Investing.com ↗ |
| 8 | Banco Santander | 3 | Model Arena and specialised banking models (Santander AI Lab with NVIDIA) — Model Arena evaluates open-weight models against banking tasks to pick the right model for each workload and is being… (use case, pilot, Oct… | Pilot | Oct 19, 2026 | Banco Santander profile → | Banco Santander ↗ |
| 9 | U.S. Bank | 2 | Multi-model, multi-vendor stack — Proprietary and open-source models balanced with models from OpenAI, Anthropic and Google, so the bank is not stuck… (use case, in production, Jul 30, 2026) | In production | Jul 30, 2026 | U.S. Bank · technology stack → | American Banker ↗ · Banking Dive ↗ |
| 10 | Goldman | 2 | OpenAI, Google, Meta, Microsoft — Models behind the GS AI Platform: GPT via Microsoft, Gemini and Llama in 2024 (section, Jul 25, 2025) | Mentioned | Jul 25, 2025 | Goldman · technology stack → | The Wall Street Journal ↗ · CNBC ↗ |
| 11 | USAA | 1 | Multi-vendor model strategy — Google Gemini used 'extensively', AWS Bedrock, IBM Watsonx Orchestrate, open-weight models such as Meta Llama watched… (use case, in production, May 13, 2025) | In production | May 13, 2025 | USAA profile → | TechTarget ↗ |
| 12 | Commonwealth Bank | 1 | Central AI platform, guardrails and token-cost control — The Chief AI Officer is accountable for a single enterprise platform for engineers and data scientists, standard… (use case, rolling out, Sep 18, 2026) | Rolling out | Sep 18, 2026 | Commonwealth Bank · technology stack → | Retail Banker International (via Yahoo Finance) ↗ |
| 13 | CIBC | 1 | Cohere (group) — Canadian open-weight models tested (section, Jul 31, 2026) | Mentioned | Jul 31, 2026 | CIBC profile → | Financial Post ↗ |
| 14 | BNY | 1 | Eliza described: about 14,000 users, built on GPT-4, Gemini and Llama (move, Aug 15, 2024) | Mentioned | Aug 15, 2024 | BNY profile → | Fortune ↗ |
Each row quotes one record from bank profiles; the “On this site” link opens the bank's own page or sub-page with every source. Banks absent from the table have nothing matching on the record this site holds.
How is this ranked?
Data: the 141 bank profiles on this site (lib/banks), every entry dated and sourced.
Rule: Entries matching Llama, Mistral, "open-source model", "open-weight" or "open source LLM". Score = matching entries; ties to banks with a matching use case in production.
Counts: 14 of 141 profiled banks qualify; 22 distinct sources cited in the table. Data checked 2026-10-28; ranking template updated 2026-10-30.
What are the caveats?
Banks absent from the table have nothing matching on the record this site holds, which is not evidence that they do nothing. Each row links the bank's own page, where every entry carries its sources.
Which bank ranks first?
BNP Paribas SA leads with 13 entries: HelloïZ virtual assistant at Hello bank! — The latest version, using generative AI (Mistral AI models) to understand and route requests with access to human… (use case, in production, May 26, 2026). Capital One is second with 7.
How many large banks name open-weight AI models?
14 of the 141 profiled banks (120 US, 21 non-US groups). Among those listed, 10 are US banks and 4 non-US groups.
What counts, and what does not?
Entries matching Llama, Mistral, "open-source model", "open-weight" or "open source LLM". Score = matching entries; ties to banks with a matching use case in production. Source: the 141 bank profiles on this site (lib/banks), every entry dated and sourced.
Is this a measure of capability?
No. Every row is a sourced fact on the public record: what the bank, a regulator, a filing or an independent outlet has put on the record. The page does not rate how well anything works.
When was this checked, and how often does it update?
The data behind the table was last checked on 2026-10-28 (bank profiles). The page recomputes whenever that dataset changes and is republished with the site; it is never edited by hand.
When any of these banks moves on AI, you'll read it the next morning.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning