Qwen

What is Qwen?

Qwen is a high-performance open-source large language model (LLM) developed and provided by Alibaba Cloud. Its Chinese name “通義千問” (Qwen) means “an all-purpose AI assistant.” Qwen supports multiple languages including Chinese, English, and Japanese, excelling in various tasks including text generation, question answering, code generation, and mathematical problem solving. Like Llama and Mistral, it’s provided as open source, allowing research institutions and companies to freely customize and use it.

In a nutshell: “A free, intelligent AI model provided by Alibaba that supports multiple languages.”

Key points:

• What it does: A language model for text generation, natural language understanding, code generation, mathematical problem solving, and text processing in multiple languages.
• Why it matters: By providing a high-performance multilingual AI model as open source, it enables non-English-speaking regions to leverage cutting-edge AI.
• Who uses it: Asian startups, companies needing multilingual support, research institutions, AI developers, and organizations requiring Japanese language processing.

Basic information

Why it matters

Traditionally, most state-of-the-art large language models were developed primarily in English. While models like GPT-4 and Claude excel in English, they often treat Asian languages, particularly Chinese and Japanese, as secondary. This resulted in Chinese and Japanese users having access to lower-quality AI services compared to English speakers.

Qwen reduced this gap by having Alibaba Cloud focus on developing high-performance models in Chinese and Asian languages. Particularly, inference performance in Chinese has improved significantly, with Qwen achieving equal or superior scores to competitors in multiple benchmarks. Furthermore, being provided as open source allows developers worldwide to contribute to improvements, enabling continuous evolution.

Key features and services

Multilingual Support Demonstrates excellent support for multiple languages and domains including Chinese, English, Japanese, code, and mathematical symbols. Particularly renowned for natural language understanding and generation in Chinese.

Multiple Model Sizes Provides models ranging from the ultra-lightweight 0.5B version to the large-scale 72B version, accommodating various computational environments. Even a small-scale version executable on smartphones and edge devices exists.

Excellent Reasoning Ability Demonstrates high performance in tasks requiring complex problem solving, mathematical reasoning, and logical thinking. Benchmark tests show favorable results compared to same-sized models from competitors.

Open Source Provision Model weights and code are completely public, allowing developers to freely customize, improve, and adapt to specific domains.

Competitors and alternatives

Llama (Meta) — The standard choice for open-source LLMs. English-centric, but Japanese support is progressing through multiple derivative models.

Mistral (Mistral AI) — An open-source LLM by a European startup. Simple and efficient, but inferior to Qwen in multilingual support.

GPT-4 (OpenAI) — Closed-source AI with state-of-the-art performance. Excellent multilingual support, but costly and not developed by a Chinese company.

Benefits and considerations

Qwen’s greatest advantage is its high performance in multiple languages, especially Chinese and Japanese. For organizations where Asian language processing is critical, more accurate and natural output is expected compared to English-centric models. Being open source enables customization and local execution with no API fees. The availability of multiple sizes allows selection based on resource environment.

Considerations include that Alibaba is a Chinese company, so organizations must carefully consider data privacy and security concerns. Compared to Meta’s Llama, the global community is smaller, and related tools and tutorials are often English-centric. Additionally, Alibaba’s support infrastructure may be more limited than Meta’s.

Related terms

• Large Language Models (LLM) — Qwen’s foundational technology. Models that learn language patterns from vast text data.
• Natural Language Processing (NLP) — AI technology that enables machines to understand and process human language.
• Fine-Tuning — The process of additional learning of pre-trained models for specific tasks.
• Multilingual Model — Machine learning models that support multiple languages.
• Open Source — Software in a form where source code is public and anyone can use, modify, and redistribute it.

Frequently asked questions

Q: Can Qwen really be used in Japanese? A: Yes. Qwen supports multiple languages including Japanese and can execute text generation, question answering, and summarization tasks in Japanese. However, completeness in Japanese may be slightly inferior compared to Japanese-specialized derivatives of Llama.

Q: Is Alibaba’s involvement a security risk? A: Concerns vary by situation. For general use, there’s no problem. However, companies handling confidential information should verify data processing regions and security infrastructure beforehand, and decide based on compliance requirements.

Q: Can Qwen run on mobile devices? A: Yes. Ultra-lightweight versions like Qwen’s 0.5B can run on smartphones and tablets. However, depending on the complexity of language processing, larger models may be needed.