Chinese tech giant Alibaba has announced the release of a new artificial intelligence model Qwen2.
It is developed by Alibaba Cloud and is the next generation of Tongyi Qianwen (Qwen). It includes Tongyi Qianwen LLM (or Qwen), Qwen-VL and Qwen-Audio.
The Qwen2 family includes a series of five models ranging in size from 0.5 to 72 billion parameters, which are trained using data from different industries in 27 languages.
Comparison of Qwen2-series AI models from Alibaba. Data: Qwen website.Queen2-72B is the most powerful model in the series, trained on 3 trillion tokens. For comparison, the Llama-2 indicator from Meta is 2 trillion tokens, Llama-3 is 15 trillion tokens.
Qwen2 can handle a long dialog context — up to 128,000 tokens, which is comparable to GPT-4o from OpenAI. The team claims that their model is superior to Meta's LLama3 in almost all of the most important synthetic tests.
Comparison of Queen2-72B with competitors. Data: Qwen website.The independent Elo Arena platform evaluates Qwen2-72B-Instruct slightly better than GPT-4-0314, but worse than Llama3 70B and GPT-4-0125- preview.
"Compared to modern open source language models, including the previously released Qwen1.5, Qwen2 outperformed most models and demonstrated competitiveness in a number of tests aimed at language comprehension, language generation, multilingualism, programming, mathematics and thinking," the Qwen team said.
Qwen2 models demonstrate a good understanding of long contexts. The Qwen2-72B-Instruct can perform information retrieval tasks anywhere without errors and passed the Needle in a Haystack test almost perfectly. Often, the performance of different models begins to deteriorate as you interact with it.
The "Needle in a haystack" test. Data: Qwen website.Earlier, Alibaba announced the release of a chatbot with artificial intelligence Tongyi Qianwen.
Instagram Facebook and Messenger platforms In April, Meta announced the launch of the free Meta AI assistant on WhatsApp, Instagram, Facebook and Messenger. It is based on the Llama 3 language model.