t-tech/T-pro-it-2.1

T-pro-it-2.1: Enhanced Russian LLM for Instruction Following and Tool Calling

T-pro-it-2.1 is an advanced 32 billion parameter language model from t-tech, built on the Qwen 3 architecture. It significantly improves upon its predecessor, T-pro-it-2.0, with a focus on robust instruction following and sophisticated tool-calling proficiency. This model is particularly optimized for Russian language processing, featuring an efficient tokenizer for faster response generation.

Key Capabilities

• Stronger Instruction Following: Demonstrates a +9 percentage point gain over T-pro-it-2.0 in adhering to complex and strict instructions.
• Advanced Tool-Calling: Achieves performance on par with Qwen3-235B-2507 in tool-calling workflows, making it highly suitable for agentic applications.
• Improved General Capabilities: Offers better comprehension and fluency across open-domain tasks, including chat and multi-step content generation.
• Efficient Russian Inference: Optimized tokenizer ensures faster response generation for Russian text.
• Long Context Support: Natively supports a context length of 32,768 tokens, with recommendations for extending it further using rope_scaling.

Training Methodology

T-pro-it-2.1 leverages an expert merging approach. After a shared Supervised Fine-Tuning (SFT) stage, three specialized experts (Instruction Following, General, and Tool-Call) are trained via Online RL alignment (GRPO) on domain-specific synthetic data. These experts are then merged using SLERP (Spherical Linear Interpolation) to create a balanced and capable unified model.

Good for

• Applications requiring precise adherence to complex instructions.
• Developing agentic systems that utilize multi-step tool-calling workflows.
• General-purpose chat and content generation in Russian.
• Scenarios demanding efficient processing of Russian text.
• Tasks benefiting from a 32,768 token context window.