T-Bank replaces a swarm of LLMs with a single 32B model and cuts query costs by 2.8–3.9x

Large enterprises that cannot send data to external clouds run into simple arithmetic: every narrow neural network built for a single task occupies scarce GPUs. T-Bank researchers proposed two ways to shrink that bill and described them at EMNLP 2026 in Budapest, an A*-rated conference. One 32B-parameter model took over the requests of more than 200 internal bank services and half of the traffic on its LLM platform. Processing became 2.8–3.9 times cheaper than on Qwen3-235B, and in some services the savings reached 9x.

One generalist instead of a swarm of specialists

The typical setup in a secured environment looks like this: one model for instruction following, another for function calling, a third for corporate scenarios. Each has to be kept in memory, scaled and updated. The T-Bank team took a different route and, for the first time, described a reproducible post-training methodology for Russian-language LLMs with agentic skills.

The pipeline has three steps. First, the team fine-tuned Qwen3-32B on a mix of corporate and open data. Then they applied reinforcement learning in three directions: instruction following, function calling and corporate tasks. This produced specialized variants, which they merged into a single model using SLERP, or spherical interpolation of weights. Merging means inference runs one network instead of three separate ones.

A compact model beats the giant

The team compared the result with Qwen3-235B, which is roughly seven times heavier. For a bank, price is not the only concern: quality on its own tasks matters too, and here the 32B model came out ahead.

• Corporate benchmarks: 69.6 points vs 65.8 for Qwen3-235B.
• Instruction following: 0.85 vs 0.83.
• Function calling: 0.79 vs 0.77.
• Query processing cost: 2.8–3.9 times lower, up to 9x in some services.
• Load: more than 200 internal services and about half of the bank's LLM platform traffic.

The gap in the metrics is small, but the direction matters. A model tuned for a specific environment does not lag behind a general-purpose giant on that environment's own tasks, while using noticeably less hardware.

Parsing PDFs on a single H100

The second paper deals with digitizing paper documents. The researchers built a pipeline for selecting difficult PDF examples and used it to train a vision-language model based on Qwen3.5-35B-A3B-Base. The model extracts any payment details and amounts from a text prompt, so there is no need to write a separate template for each document type.

The architecture is built for savings. The model has 35B parameters in total, but only 3B are activated per request. A single Nvidia H100 is enough to run it. Document parsing costs 80% less than fully manual processing and 50% less than the best open alternatives.

Note the choice of data. The key contribution is the selection of difficult examples, not simply more scans: the model is trained on the cases where standard solutions make mistakes.

What you can take into your own stack

The researchers published their methodology for building internal benchmarks, a description of the post-training pipeline, and a language model with open weights trained on public data. It is available on Hugging Face (t-tech/T-pro-it-2.1). The bank's corporate data stays in-house, but the recipe can be repeated on your own data.

Teams running closed environments can draw several conclusions. If you already have several narrow models on the same base, it makes sense to test SLERP merging after RL: it is a way to shrink your instance fleet without losing quality on key tasks. Build internal benchmarks before migrating, because that is where the 32B model showed its advantage over the 235B. For document processing, pairing an MoE with few active parameters and a single H100 means the task no longer needs a dedicated cluster and fits on one card.

For now, these are the results of one bank on its own workload, and transferring them to other industries will require separate measurements. Still, the savings figures and the open materials make the work useful for anyone who counts the cost of a token in closed infrastructure.