Mizuho’s Custom AI Matches GPT-5.2 Accuracy on Wall Street Speed, Keeps Bank Data Strictly On-Premise
Mizuho Financial Group has successfully developed a proprietary, finance-specific Large Language Model (LLM) capable of delivering highly accurate, sub-second responses without relying on prolonged AI reasoning processes.
Crucially, the new model operates entirely within Mizuho’s secure, on-premise network, allowing the bank to process highly confidential data with the sophistication of advanced general-purpose models like GPT-5.2, but without the security risks of transmitting data to external APIs.
The Speed and Security Imperative
While the adoption of generative AI has surged across the financial sector, major banks have run into a persistent bottleneck: latency. Processing complex inquiries regarding financial products, internal regulations, and risk tolerance often results in sluggish AI response times and spiraling computing costs.
In high-stakes arenas such as corporate finance, trading, and market analysis, these delays can directly translate into lost market opportunities. Furthermore, general-purpose models often struggle with the nuances of financial compliance, leading to inconsistent interpretations of institutional rules.
To bypass these hurdles, Mizuho engineered a custom model based on the Qwen3-32B open-weight architecture. By aggressively fine-tuning the AI—a process Mizuho calls "knowledge fixation"—the bank embedded fundamental financial knowledge and compliance logic directly into the model. This allows the AI to immediately output accurate answers without having to execute complex, time-consuming "reasoning" steps for every prompt.
Beating General AI on Practicality
In practical bank-level stress tests covering deposits, loans, and foreign exchange operations, Mizuho’s specialized LLM achieved an 89.0% accuracy rate with an average response time of less than one second.
By comparison, a general-purpose model like GPT-5.2 required explicit reasoning protocols to achieve a comparable accuracy of 89.7%, dragging its average response time out to a sluggish 67.4 seconds. While GPT-5.2 could theoretically match the sub-second speed and 89% accuracy if provided with the right context and stripped of its reasoning requirements, utilizing it still requires sending sensitive banking data to an external API. Mizuho’s model achieves this top-tier performance entirely behind the bank's firewall.
A Phased Rollout Toward a Multi-Agent Future
Mizuho views this breakthrough as merely the first phase of a broader technological overhaul. The bank has outlined a three-step strategy for its AI infrastructure:
- Phase 1: Finance-Specific LLM (Current): Broadly trained on financial fundamentals, laws, and internal procedures to assist with general inquiries and document creation.
- Phase 2: Domain-Specific Models: Training deeply specialized LLMs for individual departments, including Lending, Legal, and Markets, to support complex tasks like credit decision-making and drafting approval documents.
- Phase 3: Collaborative Expert LLMs: Linking multiple department-specific models into a "multi-agent system" capable of cross-departmental judgment and complex institutional decision-making.
Moving forward, Mizuho plans to continue scaling the model's parameter size, expanding its training data, and exploring advanced machine learning techniques like reinforcement learning and model merging.
As banks globally race to integrate artificial intelligence, Mizuho’s latest deployment signals a clear shift in strategy for financial titans: moving away from off-the-shelf, cloud-based AIs, and toward highly secure, hyper-specialized, on-premise infrastructure.
Norbert Gehrke
