Elamur Research⌕
☰
home / papers / 2026 / T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground

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

Paper · 2026 · Dmitrii Stoianov, Danil Taranets, Olga Tsymboi et al.

Summary

Open Russian LLMs that can reason are scarce. Most strong Russian systems are closed, and the open ones are mostly light adaptations of multilingual models. There are also few public tools for comparing fast direct answers with step-by-step reasoning or for seeing how decoding speed affects use. The T-Tech team built T-pro 2.0 on top of Qwen3-32B. They replaced rarely used non-Cyrillic vocabulary entries with Cyrillic ones and then ran an instruction-only midtraining stage on 40B tokens, with responses regenerated by a larger teacher model. After that came SFT covering both reasoning and non-reasoning modes, DPO guided by a custom reward model, and an EAGLE-style speculative decoding head to make Russian generation faster. The model can switch between quick replies and explicit reasoning traces. It segments Cyrillic text more efficiently and keeps English math and code scores close to the base model. The team also released the weights, a large SFT corpus, a Russian olympiad math benchmark and a live demo with telemetry. Together these give an open, reproducible base for research on Russian reasoning models.

Key points

  • Tokenizer: about 34k low-frequency tokens in the Qwen3 vocabulary were swapped for Cyrillic ones, with the vocabulary size kept the same.
    • On Russian Wikipedia, the share of words split into at most two tokens rose from 38% to 60%.
    • Across eight Cyrillic languages, including Ukrainian, average tokens per word fell to 2.71, against 3.63 for Qwen3.
  • Midtraining: a 40B-token mix (49% Russian, 36% English) using instruction data only raised ruAIME 2024 from 0.60 to 0.67. In 8B-scale tests, the new tokenizer gave a higher MERA macro-average than the original (0.574 vs 0.560).
  • English retention: T-pro 2.0 scores 0.765 on AIME 2024, 0.966 on MATH-500 and 0.556 on LiveCodeBench. Qwen3-32B scores 0.808, 0.961 and 0.546 on the same tests.
  • Released resources:
    • Under Apache-2.0: the model weights, the EAGLE weights and the T-Math benchmark.
    • Under ODC-By: T-Wix, a hybrid-reasoning SFT corpus of about 500k samples.
  • Demo and limitations: the demo runs on SGLang with one H100 per model. It offers side-by-side comparison with Qwen3-32B and reports token acceptance and throughput. Known limitations include offline-only alignment (DPO), long contexts beyond 32k that have not been validated, and evaluation limited to Russian and English.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Dmitrii Stoianov, Danil Taranets, Olga Tsymboi, Ramil Latypov, Almaz Dautov, Vladislav Kruglikov, Surkov Nikita, German Abramov, Pavel Gein, Dmitry Abulkhanov, Mikhail Gashkov, Viktor Zelenkovskiy, Artem Batalov, Aleksandr Medvedev and Anatolii Potapov. 2026. T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-demo.22

@inproceedings{stoianov2026t,
  title = {{T-pro 2.0: An Efficient Russian Hybrid-Reasoning Model and Playground}},
  author = {Dmitrii Stoianov, Danil Taranets, Olga Tsymboi, Ramil Latypov, Almaz Dautov, Vladislav Kruglikov, Surkov Nikita, German Abramov, Pavel Gein, Dmitry Abulkhanov, Mikhail Gashkov, Viktor Zelenkovskiy, Artem Batalov, Aleksandr Medvedev and Anatolii Potapov},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-demo.22},
  url = {https://doi.org/10.18653/v1/2026.eacl-demo.22}
}

References

71 references, 27 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2025. Turbo-alignment. https://github.com/ turbo-llm/turbo-alignment. GitHub repository. ↗
  2. Amro Abbas, Kushal Tirumala, Dániel Simig, Surya Ganguli, and Ari S. Morcos. 2023. Semdedup: Data-efficient learning at web-scale through semantic deduplication. arXiv preprint arXiv:2303.09540.
  3. Berk Atil, Sarp Aykent, Alexa Chittams, Lisheng Fu, Rebecca J. Passonneau, Evan Radcliffe, Guru Rajan Rajagopal, Adam Sloan, Tomasz Tudrej, Ferhan Ture, Zhe Wu, Lixinyu Xu, and Breck Baldwin. 2025. Non-determinism of "deterministic" llm settings. Preprint, arXiv:2408.04667. — [arXiv:2408.04667](https://arxiv.org/abs/2408.04667) find
  4. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. Preprint, arXiv:2502.13…
  5. Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shahaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, Alexander Bukharin, Yian Zhang, Tugrul Konuk, and 114 others. 2025. Llama-nemotron:…
  6. RALPH ALLAN BRADLEY and MILTON E. TERRY. 1952. Rank analysis of incomplete block designs: The method of paired comparisons. Biometrika, 39(3-4):324–345. find
  7. Peng Cui and Mrinmaya Sachan. 2025. Investigating the zone of proximal development of language models for in-context learning. Preprint, arXiv:2502.06990. find
  8. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025a. Deepseek-r1: Incentivizing reasoning capability in l…
  9. DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, and 181 others. 2025b. Deepseek-v3 technical report. Preprint, arXiv:… find
  10. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  11. Alena Fenogenova, Artem Chervyakov, Nikita Martynov, Anastasia Kozlova, Maria Tikhonova, Albina Akhmetgareeva, Anton Emelyanov, Denis Shevelev, Pavel Lebedev, Leonid Sinev, Ulyana Isaeva, Katerina Kolomeytseva, Daniil Moskovskiy, Elizaveta Goncharova, Nikita Savushkin, Polina Mikhailova, Anastasia M… find
  12. Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Rozière, and 1 others. 2024. Better & faster large language models via multi-token prediction. arXiv preprint arXiv:2404.19737.
  13. Shiyang Gu and 1 others. 2024. Sglang: Efficient serving of llms with speculative decoding and continuous batching. GitHub repository. find
  14. Igor Gusev. 2023. Saiga: Instruction-tuned russian llama models. https://huggingface.co/ IlyaGusev/saiga. ↗
  15. Dan Hendrycks and 1 others. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.
  16. Hugging Face. 2025. Open r1: A fully open reproduction of deepseek-r1. find
  17. Shawn Im and Sharon Li. 2025. Can dpo learn diverse human values? a theoretical scaling law. Preprint, arXiv:2408.03459. — [arXiv:2408.03459](https://arxiv.org/abs/2408.03459) find
  18. Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Reza Yazdani Aminadabi, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. 2024. System optimizations for enabling training of extreme long sequence transformer models. In Proceedings of the 43rd ACM Symposium on Principles of Dist… find
  19. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. 2024. Livecodebench: Holistic and contamination free evaluation of large language models for code. Preprint, arXiv:2403.07974.
  20. Diederik P. Kingma and Jimmy Ba. 2017. Adam: A method for stochastic optimization. Preprint, arXiv:1412.6980.
  21. Alexander Kukushkin. 2024. wildchat-hard-ru. https://github.com/kuk/wildchat-hard-ru. GitHub repository. ↗
  22. Ilya Kuleshov, Pavel Ilin, Nikolay Kompanets, Ksenia Sycheva, and Aleksandr Nikolich. 2025. Doom: Difficult olympiads of math. ArXiv preprint. find
  23. Yuri Kuratov and Alexey Arkhipov. 2019. Adaptation of deep bidirectional multilingual transformers for russian language. arXiv preprint arXiv:1912.11283. find
  24. Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023a. Fast inference from transformers via speculative decoding. In Proceedings of the 40th International Conference on Machine Learning (ICML ’23), pages 19274–19286. find
  25. Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023b. Fast inference from transformers via speculative decoding. arXiv preprint arXiv:2211.17192. find
  26. Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024a. From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning. Preprint, arXiv:2308.12032.
  27. Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Banghua Zhu, Joseph E Gonzalez, and Ion Stoica. 2024b. From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. arXiv preprint arXiv:2406.11939.
  28. Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Banghua Zhu, Joseph E. Gonzalez, and Ion Stoica. 2024c. From live data to high-quality benchmarks: The arena-hard pipeline. find
  29. Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024d. Eagle-2: Faster inference of language models with dynamic draft trees.
  30. Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024e. Eagle: Speculative sampling requires rethinking feature uncertainty. arXiv preprint arXiv:2401.15077.
  31. Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2025. Eagle-3: Scaling up inference acceleration of large language models via training-time test. arXiv preprint arXiv:2503.01840. — [arXiv:2503.01840](https://arxiv.org/abs/2503.01840) find
  32. Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.
  33. Chris Yuhao Liu, Liang Zeng, Yuzhen Xiao, Jujie He, Jiacai Liu, Chaojie Wang, Rui Yan, Wei Shen, Fuxiang Zhang, Jiacheng Xu, Yang Liu, and Yahui Zhou. 2025a. Skywork-reward-v2: Scaling preference data curation via human-ai synergy. arXiv preprint arXiv:2507.01352.
  34. Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. 2025b. Pairjudge rm: Perform best-of-n sampling with knockout tournament. arXiv preprint arXiv:2501.13007. In progress work. — [arXiv:2501.13007](https://arxiv.org/abs/2501.13007) find
  35. Zihe Liu, Jiashun Liu, Yancheng He, Weixun Wang, Jiaheng Liu, Ling Pan, Xinyu Hu, Shaopan Xiong, Ju Huang, Jian Hu, Shengyi Huang, Johan Obando-Ceron, Siran Yang, Jiamang Wang, Wenbo Su, and Bo Zheng. 2025c. Part i: Tricks or traps? a deep dive into rl for llm reasoning. Preprint, arXiv:2508.08221. find
  36. Keming Lu, Hongyi Yuan, Zheng Yuan, Runji Lin, Junyang Lin, Chuanqi Tan, Chang Zhou, and Jingren Zhou. 2024. \#instag: Instruction tagging for analyzing supervised fine-tuning of large language models. In The Twelfth International Conference on Learning Representations.
  37. Saumya Malik, Valentina Pyatkin, Sander Land, Jacob Morrison, Noah A. Smith, Hannaneh Hajishirzi, and Nathan Lambert. 2025. Rewardbench 2: Advancing reward model evaluation. Preprint, arXiv:2506.01937. — [arXiv:2506.01937](https://arxiv.org/abs/2506.01937) find
  38. Valentin Mamedov, Evgenii Kosarev, Gregory Leleytner, Ilya Shchuckin, Valeriy Berezovskiy, Daniil Smirnov, Dmitry Kozlov, Sergei Averkiev, Lukyanenko Ivan, Aleksandr Proshunin, Ainur Israfilova, Ivan Baskov, Artem Chervyakov, Emil Shakirov, Mikhail Kolesov, Daria Khomich, Daria Latortseva, Sergei Po… find
  39. Aleksandr Nikolich, Konstantin Korolev, Sergei Bratchikov, Igor Kiselev, and Artem Shelmanov. 2024. Vikhr: The family of open-source instruction-tuned large language models for russian. arXiv preprint arXiv:2405.13929. — [arXiv:2405.13929](https://arxiv.org/abs/2405.13929) find
  40. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M adry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card.…
  41. Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, Hanjie Chen, and Xia Hu. 2025. Stop overthinking: A survey on efficient reasoning for large language models. Preprint, arXiv:2503.16419.
  42. T-Tech. 2025a. ru-alpaca-eval. https://huggingface.co/datasets/t-tech/ Hugging Face Datasets; ru-alpaca-eval. accessed 2025. ↗
  43. T-Tech. 2025b. ru-arena-hard. https://huggingface.co/datasets/t-tech/ru-arena-hard. Hugging Face Datasets; accessed 2025. ↗
  44. OpenAI. 2025. Introducing o3 and o4-mini. https://openai.com/index/o3-o4-mini-system-card/. Accessed: 2025-05-15. ↗
  45. Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Biber. 2024. Language model tokenizers introduce unfairness between languages. arXiv preprint arXiv:2305.15425.
  46. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  47. Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.
  48. David Rein and 1 others. 2023. Gpqa: A google-proof q&a benchmark for large language models. arXiv preprint arXiv:2311.16452. find
  49. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347. find
  50. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  51. Oleh Shliazhko, Alena Fenogenova, Maria Tikhonova, Vladislav Mikhailov, Anastasia Kozlova, and Tatiana Shavrina. 2023. mgpt: Few-shot learners go multilingual. Preprint, arXiv:2204.07580.
  52. Jacob Mitchell Springer, Sachin Goyal, Kaiyue Wen, Tanishq Kumar, Xiang Yue, Sadhika Malladi, Graham Neubig, and Aditi Raghunathan. 2025. Over-trained language models are harder to fine-tune. In Forty-second International Conference on Machine Learning. find
  53. T-Tech. 2025c. ru-mt-bench. https://huggingface.co/datasets/t-tech/ru-mt-bench. Hugging Face Datasets; accessed 2025. ↗
  54. T-Tech. 2025d. ru-reasoning-benchmarks. https://huggingface.co/collections/ t-tech/ru-reasoning-benchmarks. Hugging Face Datasets; accessed 2025. ↗
  55. T-Tech. 2025e. ruaime-2024. https://huggingface.co/datasets/t-tech/ruAIME-2024. Hugging Face Datasets; accessed 2025. ↗
  56. T-Tech. 2025f. rummlu-pro. https://huggingface.co/datasets/t-tech/ruMMLU-pro. Hugging Face Datasets; accessed 2025. ↗
  57. T-Tech. 2025g. T-math. https://huggingface. Hugging Face co/datasets/t-tech/T-math. Datasets; accessed 2025. find
  58. Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Yuan Tang, Alejandro Cuadron, Chenguang Wang, Raluca Popa, and Ion Stoica. 2025. Judgebench: A benchmark for evaluating LLM-based judges. In The Thirteenth International Conference on Learning Representations.
  59. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penche… find
  60. Mikhail Tikhomirov and Daniil Chernyshev. 2024. Facilitating large language model russian adaptation with learned embedding propagation. arXiv preprint arXiv:2412.21140. — [arXiv:2412.21140](https://arxiv.org/abs/2412.21140) find
  61. Tianchun Wang, Zichuan Liu, Yuanzhou Chen, Jonathan Light, Weiyang Liu, Haifeng Chen, Xiang Zhang, and Wei Cheng. 2025a. On the effect of sampling diversity in scaling llm inference. Preprint, arXiv:2502.11027. — [arXiv:2502.11027](https://arxiv.org/abs/2502.11027) find
  62. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, and 1 others. 2023. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560. find
  63. Zhilin Wang, Jiaqi Zeng, Olivier Delalleau, Hoo-Chang Shin, Felipe Soares, Alexander Bukharin, Ellie Evans, Yi Dong, and Oleksii Kuchaiev. 2025b. HelpSteer3-Preference: Open human-annotated preference data across diverse tasks and languages. Preprint, arXiv:2505.11475. — [arXiv:2505.11475](https://arxiv.org/abs/2505.11475) find
  64. Wei Xiong, Jiarui Yao, Yuhui Xu, Bo Pang, Lei Wang, Doyen Sahoo, Junnan Li, Nan Jiang, Tong Zhang, Caiming Xiong, and Hanze Dong. 2025. A minimalist approach to llm reasoning: from rejection sampling to reinforce. Preprint, arXiv:2504.11343. — [arXiv:2504.11343](https://arxiv.org/abs/2504.11343) find
  65. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  66. An Yang, Baosong Yang, Binyuan Hui, and 1 others. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671. find
  67. Xiang Yue, Tuney Zheng, Ge Zhang, and Wenhu Chen. 2024. Mammoth2: Scaling instructions from the web. In Advances in Neural Information Processing Systems, volume 37, pages 90629–90660. Curran Associates, Inc.
  68. Yifan Zhang and Team Math-AI. 2024. American invitational mathematics examination (aime) 2024. find
  69. Yifan Zhang and Team Math-AI. 2025. American invitational mathematics examination (aime) 2025. find
  70. Weilin Zhao, Tengyu Pan, Xu Han, Yudi Zhang, Ao Sun, Yuxiang Huang, Kaihuo Zhang, Weilun Zhao, Yuxuan Li, Jie Zhou, Hao Zhou, Jianyong Wang, Zhiyuan Liu, and Maosong Sun. 2025. FR-spec: Accelerating large-vocabulary language models via frequency-ranked speculative sampling. In Proceedings of the 63r…
  71. Dmitry Zmitrovich, Alexander Abramov, Andrey Kalmykov, Maria Tikhonova, Ekaterina Taktasheva, Danil Astafurov, Mark Baushenko, Artem Snegirev, Vitalii Kadulin, Sergey Markov, Tatiana Shavrina, Vladislav Mikhailov, and Alena Fenogenova. 2023. A family of pretrained transformer language models for rus…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/stoianov-2026-t-pro-2-0-efficient/