alibaba-pai/DistilQwen2.5-DS3-0324-32B

DistilQwen2.5-DS3-0324-32B: Fast-Thinking Reasoning Model

The DistilQwen2.5-DS3-0324-32B, developed by Alibaba PAI, is a 32.8 billion parameter model built upon the Qwen2.5 base. It addresses the challenge of balancing efficient reasoning with cognitive capabilities by distilling the fast-thinking abilities of DeepSeekV3-0324 into a more lightweight architecture. This model significantly enhances reasoning speed and reduces resource consumption, making it suitable for scenarios requiring efficient on-device or edge computing.

Key Innovations & Capabilities

• Fast-Thinking Distillation Framework: Employs a two-stage process involving fast-thinking CoT (Chain-of-Thought) data collection and CoT trajectory cognitive alignment. This includes long-to-short rewriting to extract key reasoning steps and dynamic difficulty grading to optimize knowledge transfer.
• Enhanced Reasoning Speed: Achieves a 5-8x speed gain by reducing output tokens by 60-80% compared to traditional slow-thinking models, as demonstrated by metrics like MMLU_PRO and AIME2024 tokens.
• Reduced Resource Consumption: Its optimized design makes it suitable for deployment in resource-constrained environments.
• Cognitive Bias Elimination: Incorporates proprietary trajectory alignment technology to eliminate cognitive biases.
• Performance Breakthroughs: The 32B model approaches the performance of closed-source models with 10x the parameters on the GPQA Diamond benchmark.

When to Use This Model

• Efficient Reasoning Tasks: Ideal for applications where rapid and concise reasoning outputs are critical.
• Edge Computing: Its reduced resource consumption makes it well-suited for deployment on edge devices.
• Cognitive Alignment: Beneficial for tasks requiring accurate and unbiased reasoning trajectories.
• Qwen2.5 Ecosystem: Integrates seamlessly with existing Qwen2.5-based workflows due to its fine-tuning on the Qwen2.5 base model.