Unlocking the Power of Qwen2.5-Max: A Breakthrough in Large-Scale MoE Models

In the rapidly evolving landscape of artificial intelligence, scaling both data size and model complexity has proven to be a key driver of improved performance. However, effectively scaling extremely large models—whether dense or Mixture-of-Expert (MoE)—remains a challenge for researchers and industry practitioners alike. With the recent release of DeepSeek V3, many critical insights into this process have come to light. Building on these advancements, we are thrilled to introduce Qwen2.5-Max, our latest large-scale MoE model that pushes the boundaries of AI intelligence.

Pretrained on over 20 trillion tokens and further refined through curated Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), Qwen2.5-Max is designed to deliver unparalleled performance across a wide range of tasks. Today, we’re excited to share its benchmark results and announce the availability of its API via Alibaba Cloud. Let’s dive into what makes Qwen2.5-Max stand out.

Performance That Sets New Standards

To evaluate Qwen2.5-Max, we compared it against leading proprietary and open-weight models across several benchmarks that matter most to the AI community:

• MMLU-Pro: Tests knowledge with college-level problems.
• LiveCodeBench: Assesses coding capabilities.
• LiveBench: Measures general capabilities comprehensively.
• Arena-Hard: Approximates human preferences.

Our findings reveal that Qwen2.5-Max outperforms DeepSeek V3 in benchmarks like Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also delivering competitive results in other assessments such as MMLU-Pro.

When comparing base models, we benchmarked Qwen2.5-Max against top contenders like DeepSeek V3 (a leading open-weight MoE model), Llama-3.1-405B (the largest open-weight dense model), and Qwen2.5-72B (another top-tier dense model). The results speak for themselves: our base models show significant advantages across most benchmarks.

These achievements underscore not only the scalability of MoE architectures but also the potential of advanced post-training techniques to elevate future iterations of Qwen2.5-Max even further.

How to Get Started with Qwen2.5-Max

Excited to try Qwen2.5-Max? You can now interact with it directly through Qwen Chat, where you can chat with the model, explore artifacts, conduct searches, and more.

For developers looking to integrate Qwen2.5-Max into their applications, the API is available under the name qwen-max-2025-01-25. Here’s how you can get started:

1. Register an account on Alibaba Cloud and activate the Model Studio service.
2. Navigate to the console and create an API key.
3. Use the API, which is fully compatible with OpenAI-API standards, in your projects.

Below is a simple Python example to help you get started:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

completion = client.chat.completions.create(
    model="qwen-max-2025-01-25",
    messages=[
        {'role': 'system', 'content': 'You are a helpful assistant.'},
        {'role': 'user', 'content': 'Which number is larger, 9.11 or 9.8?'}
    ]
)

print(completion.choices[0].message)

With just a few lines of code, you can harness the power of Qwen2.5-Max for your downstream applications, including chatbots, coding assistants, and more.

The Road Ahead

The journey toward building ever-smarter AI systems continues. At Qwen, we believe that scaling data and model size is not just about improving performance—it’s about advancing the frontiers of human knowledge. Our focus remains on enhancing the reasoning and thinking capabilities of large language models through innovative reinforcement learning techniques.

This endeavor holds immense promise: enabling models to surpass human-level intelligence and unlock new realms of understanding. As we look to the future, we invite researchers, developers, and enthusiasts to join us in exploring the untapped potential of AI.

Cite Us!

If you find Qwen2.5-Max useful in your work, please feel free to cite our technical report:

@article{qwen25,
  title={Qwen2.5 technical report},
  author={Qwen Team},
  journal={arXiv preprint arXiv:2412.15115},
  year={2024}
}

Join the Conversation

We’re eager to hear your thoughts and experiences with Qwen2.5-Max! Share your feedback, ask questions, and connect with fellow users on our Discord server. Together, let’s shape the future of AI.

Explore Qwen2.5-Max today at Qwen Chat and unleash the next generation of intelligent applications!