The Underrated Qwen

Originally published in Chinese on March 2, 2025. Qwen is Alibaba's family of foundation models. Here is what it had released by early March 2025. Qwen 2.5-Max Core features: built on a very large MoE (mixture of experts) architecture, pretrained on 20 trillion tokens, supporting more than 29 languages including Chinese, English, and Japanese. It supports a 128K context length and can generate 8K of content, suitable for long-text processing and complex tasks such as report generation. It has multimodal capability (visual understanding, video analysis) and coding assistance, and can generate code and optimize logic. Performance: it beats DeepSeek V3 and Llama-3.1-405B on benchmarks such as Arena-Hard and LiveBench, standing out in language understanding and generation. The "deep thinking" (QwQ) reasoning model: a reasoning model optimized from Qwen2.5-Max, specialized in mathematical understanding, coding, and agent interaction, with web search support. It shows the full chain of thought, suited to scenarios that need step-by-step reasoning, such as solving math problems. It is currently a preview (QwQ-Max-Preview). A full release and apps for Android and iOS are planned, with open weights. Lightweight versions such as QwQ-32B are offered for local deployment. The Qwen2.5-VL multimodal series Sizes: 3B, 7B, and 72B, covering both on-device deployment (3B) and high performance (72B). Core capabilities: dynamic-resolution visual encoding that strengthens image and video understanding, with especially strong table parsing that surpasses GPT-4o and Claude 3.5. It has spatial and temporal awareness, can handle one hour of video, and supports structured outputs such as coordinates and JSON. The license is Apache 2.0, free for developers to use commercially.