The Year in AI Papers
The year/Independent research

Paper 2512.10430

T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground

Dmitrii Stoianov, Danil Taranets, Olga Tsymboi, Ramil Latypov, Almaz Dautov, Vladislav Kruglikov, Nikita Surkov, German Abramov, Pavel Gein, Dmitry Abulkhanov, Mikhail Gashkov, Viktor Zelenkovskiy, et al.

Published
Dec 2025
Research lab
Independent
Citations
1
GitHub
Not linked

01 In brief

Summary

T-pro 2.0 is an open-weight Russian LLM for hybrid reasoning and efficient inference, supporting direct answering and reasoning-trace generation.

It uses a Cyrillic-dense tokenizer and an adapted EAGLE speculative-decoding pipeline to reduce latency.

The authors release the model weights, the T-Wix 500k instruction corpus, the T-Math reasoning benchmark, and EAGLE weights on Hugging Face, along with a public web demo for side-by-side comparison of reasoning and non-reasoning modes.

Training includes instructional midtraining on 40B tokens, SFT, and on-policy DPO.

The model achieves strong results on Russian benchmarks (MERA 0.66, ruMMLU-Pro 0.697, Arena Hard Ru 91.1) and competitive English performance.

EAGLE decoding yields an average 1.85x speedup.

T-Math, a new benchmark of 331 olympiad problems, shows T-pro 2.0 at 0.541 pass@1.

The model is released under Apache-2.0, with T-Wix under ODC-By.

Limitations include no agentic capabilities, offline-only RL, and unverified long-context performance beyond 32k tokens.

02 From the paper

Abstract

We introduce T-pro 2.0, an open-weight Russian LLM for hybrid reasoning and efficient inference. The model supports direct answering and reasoning-trace generation, using a Cyrillic-dense tokenizer and an adapted EAGLE speculative-decoding pipeline to reduce latency. To enable reproducible and extensible research, we release the model weights, the T-Wix 500k instruction corpus, the T-Math reasoning benchmark, and the EAGLE weights on Hugging Face. These resources allow users to study Russian-language reasoning and to extend or adapt both the model and the inference pipeline. A public web demo exposes reasoning and non-reasoning modes and illustrates the speedups achieved by our inference stack across domains. T-pro 2.0 thus serves as an accessible open system for building and evaluating efficient, practical Russian LLM applications.