T-Bank's AI Center has introduced two training methods designed to reduce costs for companies running AI models on their own infrastructure. The first approach replaces multiple specialized language models with a single universal model. By utilizing a 32-billion parameter model based on Qwen3-32B, the bank consolidated over 200 internal services, handling half of its internal LLM platform traffic monthly.
This universal model processes requests 2.8 to 3.9 times cheaper than the 235-billion parameter Qwen3-235B, with cost reductions reaching 4 to 9 times in specific services. In benchmarks, the model scored 69.6 points for corporate queries, 0.85 for instruction following, and 0.79 for function calling, outperforming larger models. The training process involves reinforcement learning via GRPO and weight merging using SLERP.
The second method focuses on visual-language models for document processing. Using a 35-billion parameter model based on Qwen3.5-35B-A3B-Base, which activates only 3 billion parameters per request, the bank achieved significant efficiency. Fine-tuning this model requires only one Nvidia H100 GPU. The approach reduces document processing costs by over 80% compared to manual labor and over 50% compared to existing open-source models.
To support these methods, researchers developed the Difficulty-Aware Data Curation pipeline for document training and an economic efficiency model to balance automation with human intervention. T-Bank has released the methodology for creating internal benchmarks, the post-training approach, and a version of the language model with open weights for external use.
