Qwen2.5-Max Release Notes

Qwen has announced the release of Qwen2.5-Max, a large-scale Mixture-of-Experts (MoE) model designed to scale intelligence through massive data and model size. The model is now available for use via the Qwen Chat interface and through an OpenAI-compatible API provided by Alibaba Cloud.

Model Architecture and Training

Qwen2.5-Max is built on a Mixture-of-Experts (MoE) architecture. It was pretrained on a dataset exceeding 20 trillion tokens. To refine its capabilities, the model underwent post-training using curated Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) methodologies.

Performance Benchmarks

Qwen2.5-Max demonstrates competitive performance against both proprietary and open-weight models across several key benchmarks:

Instruct Model Performance

When compared to state-of-the-art models including GPT-4o, Claude-3.5-Sonnet, and DeepSeek V3, the Qwen2.5-Max instruct model outperforms DeepSeek V3 in the following areas:

  • Arena-Hard
  • LiveBench
  • LiveCodeBench
  • GPQA-Diamond

It also shows competitive results in MMLU-Pro, which tests college-level knowledge.

Base Model Performance

In evaluations of base models, Qwen2.5-Max was compared against Llama-3.1-405B, Qwen2.5-72B, and DeepSeek V3. The Qwen team reports that the Qwen2.5-Max base model demonstrates significant advantages across most of these benchmarks.

Availability and Integration

Qwen2.5-Max (model name qwen-max-2025-01-25) is accessible through two primary channels:

  1. Qwen Chat: Users can interact with the model directly, utilizing features such as artifacts and search.
  2. Alibaba Cloud API: The API is OpenAI-compatible, allowing developers to integrate the model using standard OpenAI Python libraries by pointing the base_url to https://dashscope-intl.aliyuncs.com/compatible-mode/v1.

Future Research Directions

Qwen is focusing on the application of scaled reinforcement learning to further enhance the thinking and reasoning capabilities of large language models, with the goal of enabling models to move beyond human-level intelligence.

Sources