Qwen 2 Instruct (72B)

State-of-the-art AI model improving customer engagements through sophisticated chat functionalities.

Introduction

Qwen 2 is out, and it's stronger than the last generation of Qwen 1.5. The model's linguistic proficiency has been broadened to 27 additional languages, demonstrated state-of-the-art results across a multitude of evaluations, and the context length support was increased up to an impressive 128K tokens. This enhancement allows for more comprehensive and contextually rich interactions, making Qwen2 an even more powerful tool for a variety of applications. Qwen2 builds on the Transformer architecture, adding advanced features like SwiGLU activation, attention QKV bias, group query attention, a mixture of sliding window attention, and more for improved efficiency and focus when processing information.

The Model

A notable enhancement over its predecessor, Qwen1.5, is the universal application of Group Query Attention, ensuring faster speeds and reduced memory usage during inference. Context length capabilities have been significantly improved to 128K, and strongly tested to have no blind spots on big context inputs. This model is not limited to English and Chinese, it includes spanning regions from Western Europe to Southern Asia and includes a robust approach to handling code-switching. This multilingual competency is bolstered by extensive pretraining and instruction-tuning datasets, affirming the Qwen2 series as a versatile and powerful tool for diverse linguistic and contextual challenges. The model also has improved performance in coding, math, and reasoning.

Use Cases for the Model Qwen 2 Chat

Text Generation

Qwen 2 72B Instruct is well-suited for building chatbots due to its large size and impressive language generation capabilities. It can engage in meaningful and contextually relevant conversations across a variety of topics, providing a natural and human-like interaction experience.

Retrieval Augmented Generation and Function Calling

Qwen shines in tasks that need augmenting with retrieval-generation. Qwen 2 is good in the way that it can give reliable output with fewer hallucinations. Plus, it excels at using external tools (like a function call), making it ideal for building AI agents needing to interact with the real world.

Content Moderation

Qwen 2 72B Instruct excels at content moderation. Its powerful language understanding tackles inappropriate content across languages, effectively filtering harmful material. Studies show it competes with the models having the best filter (like GPT-4) and outperforms others (Mistral-8x22B) in multilingual safety tasks.

Multilingual Applications

Qwen 2 72B Instruct excels in multilingual tasks (French, Spanish, Japanese, etc.), proven across 12 languages. YARN allows it to handle complex contexts (up to 128k tokens), crucial for instruction-tuned models. To improve multilingual capabilities, it's trained in 27+ languages (Western European, Eastern European, Middle Eastern, Eastern Asian, South-Eastern Asian, Southern Asian) and tackles code-switching effectively. In our testing (LLama 3 vs Qwen 2) we took it through some not-so-obvious prompts with cultural context, like translating an idiom, and it performs reasonably well. Moreover, it is a must when working with Asian groups of languages.

How It Compares to Competitors

Qwen 2 72B Instruct t is benchmarked against competitors - Llama 3 70B, GPT-4, Mistral and othersla. The performance is higher than LLama on all classic benchmarks like GSM8K, MATH etc. GPT-4 and MIstral were compared in to Qwen 2 on fraud and pornography detection in inputs - and the results are comparable. Moreover, this AI Model has increased alignment with human preference (on the AlignBench benchmark). It also gets a large context window of 128K.

Tips

You can utilize the model for your application by signing up for AI/ML API access on this website.

To try Qwen 2 72B Instruct locally, grab the newest Huggingface Transformers library (at least version 4.37.0). Keep an eye on the generation_config.json file for settings to tackle issues like code-switching (more details in the model's repository).

License Agreement

Qwen 2-72B Instruct uses the Tongyi Qianwen license. You can find it on the model's GitHub or Hugging Face repository. For commercial use, it's free, but if your product or service has a massive user base (over 100 million monthly active users), you will need to contact the developers to request permission.

Conclusion

Qwen 2 surpasses its predecessor in all aspects. It boasts multilingual capabilities across 27 languages, tackles complex tasks with 128k context support, and excels at diverse applications like chatbot development, content moderation, and multilingual tasks. Benchmarking reveals strong performance compared to competitors, making Qwen 2 a valuable tool for various AI needs. Its user-friendly access via APIs and local installation options further enhance its accessibility.