Alibaba unveils Qwen2.5-Max, claiming wins over DeepSeek V3

Background

Alibaba's Qwen team chose DeepSeek V3 as the main comparison for its new large model.

What happened

The Qwen team announced Qwen2.5-Max in a blog post dated 28 January 2025. Wikipedia gives the launch date as 29 January 2025.^[1][2]

The model is a large-scale mixture-of-experts model pre-trained on more than 20 trillion tokens. It was post-trained with supervised fine-tuning and reinforcement learning from human feedback.^[1]

Key details

The team reported that its instruct model outperformed DeepSeek V3 on Arena-Hard, LiveBench, LiveCodeBench and GPQA-Diamond, and was competitive with it on MMLU-Pro.^[1]

Developers could reach the model through Qwen Chat and the Alibaba Cloud API, under the name qwen-max-2025-01-25. The API uses the OpenAI API format.^[1]

Alibaba did not release the model's weights.^[1]

Several terms in the announcement may be unfamiliar. An instruct model is a version that has been tuned, here with supervised fine-tuning and reinforcement learning from human feedback, so that it follows instructions and holds a conversation. Benchmarks are standard test sets used to compare models. GPQA-Diamond is a set of difficult science questions, LiveCodeBench tests models on programming problems, and Arena-Hard and LiveBench are evaluation sets built to measure general ability.

Pre-training is the initial training on a very large body of text, before any tuning. The announced model was pre-trained on more than 20 trillion tokens. A token is the unit of text that a language model reads and writes, roughly a word or part of a word. The mixture-of-experts design splits the network into many expert sub-networks and activates only some of them for each token.^[1]

Open weights means the trained parameters are published so that anyone can download and run the model. Because the weights of this model were not released, outside users could reach it only through Alibaba's own services. The OpenAI API format is a widely used request structure. Software written for it can often be pointed at another provider by changing the address and model name, so compatibility with the format makes adoption easier for developers.^[1]

Impact

Qwen2.5-Max was offered only as a hosted service and was presented against DeepSeek V3 on public benchmarks. In September 2025 Alibaba's proprietary flagship role passed to Qwen3-Max.^[2]