About Qwen
Alibaba Cloud provides Tongyi Qianwen (Qwen) model series to the open-source community. This series includes Qwen, the large language model (LLM); Qwen-VL, the large language vision model; Qwen-Audio, the large language audio model; Qwen-Coder, the coding model; Qwen-Math, the mathematical model; and QwQ-32B, the reasoning model. You can try Qwen models and easily customize and deploy them in Alibaba Cloud Model Studio.
The latest Qwen 2.5 models are pre-trained on our latest large-scale dataset, which includes up to 18 trillion tokens. Compared to Qwen2, Qwen2.5 has acquired significantly more knowledge (MMLU: 85+) and has greatly improved capabilities in coding (HumanEval 85+) and mathematics (MATH 80+). Additionally, the new models have significantly improved in following instructions, generating long texts, understanding structured data, and generating structured outputs. Qwen2.5 models are generally more resilient to the diversity of system prompts, enhancing role-play implementation and condition-setting for chatbots. Qwen2.5-Max, the large-scale MoE model, has been pretrained on over 20 trillion tokens and further post-trained with curated Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) methodologies. Qwen2.5-Coder has been trained on 5.5 trillion tokens of code-related data, delivering competitive performance against larger language models on coding evaluation benchmarks. Qwen2.5-Math supports both Chinese and English and incorporates various reasoning methods, including Chain-of-Thought (CoT), Program-of-Thought (PoT), and Tool-Integrated Reasoning (TIR). QwQ-32B leverages reinforcement learning to excel in complex problem-solving tasks like mathematical reasoning and coding, achieving performance comparable to larger models.
Leading Performance in Multiple Dimensions
Qwen outperforms other open-source baseline models of similar sizes on a series of benchmark datasets that evaluate natural language understanding, mathematical problem-solving, coding, etc.
Easy and Low-Cost Customization
You can deploy Qwen models with a few clicks in PAI-EAS, and fine-tune them with your data stored on Alibaba Cloud, or external sources, to perform industry or enterprise-specific tasks.
Applications for Generative AI Era
You can leverage Qwen APIs to build generative AI applications for a broad range of scenarios such as writing, image generation, audio analysis, etc. to improve work efficiency in your organization and transform customer experience.
Qwen Model Family
Qwen
Our latest Qwen 2.5 models have been pre-trained with high-quality data from a wide range of domains and languages, supporting a context length of up to 128K tokens. These models offer enhanced performance in coding, mathematics, human preference, and other core capabilities such as following instructions and understanding or generating structured data.
Qwen2.5-Max
Qwen2.5-Max is a large-scale Mixture-of-Expert (MoE) model that has been pretrained on over 20 trillion tokens. It demonstrated leading performance in benchmarks such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, compared to models such as DeepSeek V3 and Llama 3.1.
Qwen2.5-Coder
Qwen2.5-Coder is an open-source coding model. It supports up to 128K tokens of context, covers 92 programming languages, and has achieved remarkable improvements across various code-related evaluation tasks, including code generation, multi-programming code generation, code completion, and code repair.
Qwen2.5-Math
Qwen2.5-Math is our mathematical LLM pre-trained and fine-tuned with synthesized mathematical data. It supports bilingual queries in English and Chinese and excels in Chain-of-Thought (CoT) and Tool-Integrated Reasoning (TIR). Qwen2.5-Math outperforms most 70B math models in various tasks.
Qwen-VL
Qwen-VL is the large vision language model of the Qwen series. It generates content based on images, text, and bounding boxes as input. With leading performance verified by multiple evaluation benchmarks, Qwen-VL can perform fine-grained text recognition in both Chinese and English, compare and analyze these images, then create stories, solve math problems, or answer questions.
Qwen-Audio
Qwen-Audio is the large audio language model of the Qwen series. Qwen-Audio accepts text and diverse audio files (human speech, natural sound, music, and songs) as inputs, and provides text-based output. Qwen-Audio achieves impressive performance without any task-specific fine-tuning on the test set of Aishell1, cochlscene, ClothoAQA, and VocalSound.
QwQ-32B
QwQ-32B scales Reinforcement Learning (RL) to enhance performance and integrates agent capabilities for critical thinking and adaptive reasoning. With only 32 billion parameters, it matches the performance of DeepSeek-R1 (671B parameters). QwQ-32B is open-weight in Hugging Face and ModelScope under the Apache 2.0 license.
What Qwen Can Do
Chatbot based on Qwen, Qwen-Audio, and Qwen-VL answering questions containing multimodal data
Qwen-VL producing a cartoon-style human portrait based on prompts
Qwen-VL recognizing the objects (the woman and the dog) in the image and their gestures (high five)
Qwen-Audio analyzing the identity and emotions of the speaker and recommending replies
Qwen2.5-72B providing formatted output based on the requirement and input data (table in JSON format)
Qwen2.5-72B generating JSON code step by step with explanations as requestedQwen2.5-72B writing a report of over 5,000 Chinese characters on the requested subject
Try Qwen Models on Alibaba Cloud Model Studio
Qwen on Open-Source Communities
Hugging Face is an open-source community advancing AI collaboration through tools like the Transformers library, enabling easy access to pre-trained NLP, vision, and generative models.
ModelScope, developed by Alibaba, is an open-source platform offering diverse AI models (NLP, CV, multimodal, etc.) and tools to streamline model development, deployment, and sharing.
GitHub is the world’s largest open-source hub for code collaboration, providing version control, issue tracking, and community-driven development across software and AI projects.
