What is ACE-Bank?

ACE-Bank is an AI model developed by APMIC, purpose-built for Taiwan's financial wealth-management advisory scenarios. Built on the latest Gemma-4-26B-A4B architecture, the model has deep knowledge of local financial regulations and market logic, and has passed 13 categories of financial advisor certification exams administered by the Taiwan Academy of Banking and Finance, enabling it to provide precise, compliant wealth management advice and business support.

Technically, ACE-Bank employs an advanced hybrid attention mechanism and long-context optimization technology (p-RoPE), allowing it to process lengthy and complex regulatory texts and financial reports while maintaining the low latency and low memory consumption advantages of a lightweight model.

How was this benchmark conducted?

This benchmark pitted 6 AI models against each other: the flagship gemini-3.1-pro, ACE-Bank (ACE-3-Bank), Muse-Glimmer-30B, ACE-Bank's unfine-tuned base model gemma-4-26B-A4B-it, gpt-oss-120b, and gpt-5.6-sol. Answer options for each question were randomly reordered during testing to reduce interference from models memorizing option positions, and scoring was based on parsing model answers in \boxed{option} format. Any response from which an A/B/C/D option could not be parsed was scored as incorrect in the raw scoring.

How did ACE-Bank perform?

All 13 financial certifications passed

First, ACE-Bank's scores across all 13 subjects exceeded the 70% passing threshold, demonstrating comprehensive competency across the financial advisory domain.

The 13 subjects include: Bank Internal Control and Internal Audit Test (Consumer Finance), Financial Technology Competency, Advanced Credit Extension, Family Trust Planning Advisor Certification, Bank Internal Control and Internal Audit Test (General Finance), Junior Credit Officer, Basic Risk Management Competency, Senior/Elderly Financial Planning Advisor Certification, Trust Business Professional Test for Trust Industry Personnel, Derivatives, Sustainable Development Basic Competency Test, Junior Foreign Exchange Officer, and Financial Planner.

Significant fine-tuning gains: 11.34% improvement over base model

Compared to the pre-fine-tuned gemma-4-26B-A4B-it, whose accuracy rate was only 74.71%, the fine-tuned model improved to 83.18% — a gain of 11.34 percentage points. This confirms that APMIC's domain-specific fine-tuning for financial advisory scenarios delivers a substantial accuracy improvement.

Leading same-tier models, closely trailing flagship models

In this benchmark, the accuracy rates of the six models were: gemini-3.1-pro (92.3%), gpt-5.6-sol (91.2%), ACE-Bank (81.7%), Muse-Glimmer-30B (72.9%), base model gemma-4-26B-A4B-it (71.5%), and gpt-oss-120b (68.0%). ACE-Bank ranked third, leading its same-tier lightweight models by roughly 9 to 14 percentage points, while trailing gemini-3.1-pro and gpt-5.6-sol by 9.5 to 10.6 percentage points.

Lowest response latency, fastest response speed

ACE-Bank's average per-question latency is approximately 2.7 seconds, the fastest among the 6 models — outperforming gemma-4-26B-A4B-it (3.8 seconds), gpt-5.6-sol (5.7 seconds), gpt-oss-120b (8.9 seconds), gemini-3.1-pro (14.2 seconds), and Muse-Glimmer-30B (23.6 seconds). At a comparable level of accuracy, ACE-Bank completed all 1,346 questions in 368 seconds, equivalent to approximately 219.5 questions per minute — the fastest of all models tested.

Cost advantage: second-lowest average output tokens overall

The number of output tokens reflects the computational cost of each call, a key consideration for enterprise-scale deployment. ACE-Bank uses only about 360 tokens on average per question to complete its reasoning and answer — nearly a quarter of the tokens used by gemini-3.1-pro (1,392) and Muse-Glimmer-30B (1,463), and only slightly more than gpt-5.6-sol (309).

Summary: ACE-Bank's Competitive Advantages

This benchmark demonstrates ACE-Bank's high efficiency and low cost:

• All 13 subjects passed: All 13 subjects exceeded the 70% passing threshold, with an 11.34 percentage-point improvement over its own base model — showing that the fine-tuning not only improved performance but reached a level capable of passing certification exams.

• Third-place accuracy ranking: While its accuracy rate trails slightly behind flagship models gemini-3.1-pro and gpt-5.6-sol, it clearly leads the three same-tier lightweight models.

• Fastest response, concise output: An average per-question latency of 2.7 seconds and an average response length of about 360 tokens allow ACE-Bank to deliver over 80% accuracy at a lower time cost.

• Low computational cost: The average output tokens per question is the second-lowest among the 6 models — only about a quarter of the response length of gemini-3.1-pro and Muse-Glimmer-30B — representing a significant cost difference at scale.

ACE-Bank delivers near-top-tier accuracy and faster responses at a lower computational cost. For financial application scenarios requiring large-scale deployment, real-time performance, and cost sensitivity, ACE-Bank offers practical deployment value that balances efficiency, cost, and accuracy. If you have any needs or questions regarding this model, please feel free to contact us.