YiXin-AILab

YiXin-Agentic-Qwen3-14B: A High-Performance Agentic Model with Multi-Turn Interactions. YiXin-Agentic-Qwen3-14B features powerful reasoning and agent capabilities, developed through multi-stage distillation and reinforcement training.This model outperforms models of the same scale in all aspects of capability. Moreover, it is also highly competitive in terms of agent capabilities even when compared to large models such as DeepSeek-V3.1, Kimi-K2-Instruct-0905, and Qwen3-235B-A22B-Thinking-2507. Specifically, the training process encompasses the following four innovative design aspects. - Mulit Train Stage: Initial stage uses process-supervised knowledge distillation to boost reasoning and generalization. Later, multi-stage reinforcement learning enhances agent decision quality and stability. - More Efficient Algorithm: Adopt advanced GRPO++ algorithm. Its innovations simplify training, reduce hyperparameter tuning complexity, and improve optimization efficiency via reward-focused updates. - More Stable Training: Remove entropy loss to lower training collapse risks. Add monitoring metrics (reasoning repetition rate, truncation rate, Channel Reward). - More Diverse Exploration: Apply FIRE Sampling Rollout globally, combined with Clip High. Expands training data breadth for diverse solutions to complex/rare samples. YiXin-Agentic-Qwen3-14B was benchmarked against multiple models, including DeepSeek-V3.1, Kimi-K2-Instruct-0905, and Qwen3-235B-A22B-Thinking-2507, Qwen3-14B. The model has achieved significant improvements in both agentic and reasoning capabilities, including tool usage, logical reasoning, mathematics and coding: | Metric | DeepSeek-V3.1 | Kimi-K2-Instruct-0905 | Qwen3-235B-A22B-Thinking-2507 | Qwen3-14B | YiXin-Agentic-Qwen3-14B | |---------------|-------------|-----------------------------|------------------------------|-------------|------------------------| | TAU1-Retail | 71.3 | 54.7 | 62.0 | 49.3 | 64.1 | | TAU1-Airline | 49.3 | 63.2 | 51.3 | 26.7 | 38.7 | | TAU2-Retail | 70.4 | 40.4 | 62.9 | 42.5 | 63.2 | | TAU2-Airline | 43.0 | 50.0 | 46.0 | 32.5 | 48.0 | | TAU2-Telecom | 41.2 | 74.1 | 53.5 | 30.5 | 59.7 | | C3-Bench UA | 55.3 | 52.5 | 62.3 | 64.3 | 66.2 | | C3-Bench UPTA | 61.1 | 63.8 | 64.1 | 65.8 | 68.0 | | AutoFin | 46.9 | 38.8 | 48.1 | 33.1 | 49.5 | | AVG | 54.8 | 54.7 | 56.3 | 43.1 | 57.2 | | Metric | DeepSeek-V3.1 | Kimi-K2-Instruct-0905 | Qwen3-235B-A22B-Thinking-2507 | Qwen3-14B | YiXin-Agentic-Qwen3-14B | |---------------|-------------|-----------------------------|------------------------------|-------------|------------------------| | MATH-500 | 95.0 | 94.6 | 96 | 94.4 | 94.8 | | GPQA-Diamond | 84.3 | 77.8 | 80.8 | 65.2 | 68.7 | | AIME-24 | 90 | 80 | 90 | 76.67 | 83.3 | | AIME-25 | 80 | 70 | 83.33 | 70 | 80 | | LiveCodeBench v6 [24.08-25.05] | 74.8 | 53.7 | 74.5 | 56.6 | 57.3 | | AVG | 84.8 | 75.2 | 84.9 | 72.6 | 76.8 | Note: Since the examples involve sensitive company business information, the given prompts are simplified samples based on real scenarios from Yixin Auto Finance. For instance, you can easily start a service using vLLM: If you find our work helpful, feel free to give us a cite.