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

---

license: apache-2.0

📖 Introduction

DistilQwen2.5-DS3-0324 Series: Fast-Thinking Reasoning Models

Overview

In response to the industry challenge of balancing efficient reasoning with cognitive capabilities, the DistilQwen2.5-DS3-0324 series innovatively transfers the fast-thinking capabilities of DeepSeekV3-0324 to lightweight models. Through a two-stage distillation framework, this series achieves high performance while delivering:

• Enhanced Reasoning Speed: Reduces output tokens by 60-80% (compared to slow-thinking models)
• Reduced Resource Consumption: Suitable for edge computing deployment
• Elimination of Cognitive Bias: Proprietary trajectory alignment technology

Core Innovations

1. Fast-Thinking Distillation Framework

• Stage 1: Fast-Thinking CoT Data Collection

  • Long-to-Short Rewriting: Extracts key reasoning steps from DeepSeek-R1
  • Teacher Model Distillation: Captures the rapid reasoning trajectories of DeepSeekV3-0324

• Stage 2: CoT Trajectory Cognitive Alignment

  • Dynamic Difficulty Grading (Easy/Medium/Hard)
    • LLM-as-a-Judge evaluates small model comprehensibility
    • Simple chain expansion → Adds necessary steps
    • Hard chain simplification → Removes high-level logical leaps
  • Validation Mechanism: Iterative optimization until all data reaches "Medium" rating

2. Performance Breakthroughs

• 32B Model approaches the performance of closed-source models with 10x the parameters on the GPQA Diamond benchmark
• Significant Improvement in Reasoning Efficiency (see comparison table below)

Model	MMLU_PRO Tokens	AIME2024 Tokens	Speed Gain
DistilQwen2.5-R1-32B (Slow-Thinking)	4198	12178	1x
DistilQwen2.5-DS3-0324-32B	690	4177	5-8x

Technical Advantages

• Two-Stage Distillation: First compresses reasoning length, then aligns cognitive trajectories
• Dynamic Data Optimization: Adaptive difficulty adjustment ensures knowledge transferability
• Open-Source Compatibility: Fine-tuned based on the Qwen2.5 base model

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    "alibaba-pai/DistilQwen2.5-DS3-0324-32B",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("alibaba-pai/DistilQwen2.5-DS3-0324-32B")

prompt = "Give me a short introduction to large language model."
messages=[
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You should think step-by-step."},
    {"role": "user", "content": prompt},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=2048，
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]