Hugging Face Overview
T-pro-it-2.0: A Reasoning-Optimized Qwen 3 Model
T-pro-it-2.0 is a 32 billion parameter language model from t-tech, based on the Qwen 3 family. It has undergone extensive continual pre-training and alignment, with a particular emphasis on enhancing its reasoning capabilities. The model was trained on 40 billion tokens of instruction data, with one-third dedicated to reasoning tasks, and further fine-tuned on 500K high-quality instructions, where reasoning tasks constituted 20% of the dataset. Additionally, it benefited from preference tuning on 100K carefully selected instructions.
Key Capabilities & Differentiators
- Enhanced Reasoning: Significantly improved performance on reasoning-focused benchmarks like MERA, Ru Arena Hard, and ru AIME 2025, outperforming the base Qwen 3 32B model and other competitors.
- Flexible Reasoning Modes: Supports dynamic switching between 'thinking' and 'non-thinking' modes via the
enable_thinkingflag intokenizer.apply_chat_template, allowing for optimized responses based on task complexity. - Optimized for Russian Language: Demonstrates strong performance across various Russian language benchmarks (ruMMLU, Ru Arena Hard, ru LCB).
- Long Context Support: Natively supports a context length of 32,768 tokens, with recommendations for extending to 128K for extremely long texts.
Recommended Use Cases
- Complex Problem Solving: Ideal for tasks requiring step-by-step reasoning, mathematical calculations, and logical deduction.
- Instruction Following: Excels in adhering to diverse and complex instructions, particularly in a conversational context.
- Russian Language Applications: Well-suited for applications targeting Russian-speaking users, given its strong benchmark results in this domain.
- Interactive AI Assistants: Can be effectively deployed as a virtual assistant, especially when precise and reasoned responses are critical.