Alibaba Unveils New AI Model Qwen2

Chinese tech giant Alibaba announced the release of its new artificial intelligence model, Qwen2. 

Developed by Alibaba Cloud, it is the next generation of Tongyi Qianwen (Qwen). It includes Tongyi Qianwen LLM (or Qwen), Qwen-VL, and Qwen-Audio.

The Qwen2 family comprises a series of five models ranging from 0.5 to 72 billion parameters, trained using data from various industries in 27 languages.

Queen2-72B is the most powerful model in the series, trained on 3 trillion tokens. For comparison, Meta’s Llama-2 is trained on 2 trillion tokens, while Llama-3 is trained on 15 trillion tokens.

Qwen2 can handle long conversational contexts—up to 128,000 tokens, comparable to OpenAI’s GPT-4o. The team claims their model surpasses Meta’s LLama3 in nearly all major synthetic tests.

Independent platform Elo Arena rates Qwen2-72B-Instruct slightly better than GPT-4-0314, but lower than Llama3 70B and GPT-4-0125-preview. 

“Compared to modern open-source language models, including the previously released Qwen1.5, Qwen2 has outperformed most models and demonstrated competitiveness in a range of tests targeting language understanding, language generation, multilingualism, programming, mathematics, and reasoning”, stated the Qwen team.

The Qwen2 models exhibit strong comprehension of long contexts. Qwen2-72B-Instruct can flawlessly perform information retrieval tasks anywhere and nearly aced the “Needle in a Haystack” test. Often, the performance of different models begins to degrade with continued interaction.

Earlier, Alibaba announced the release of the AI chatbot Tongyi Qianwen. 

Back in April, Meta announced the launch of a free AI assistant Meta AI on platforms WhatsApp, Instagram, Facebook, and Messenger. It is based on the Llama 3 language model.