Alibaba Cloud, the cloud computing arm of Alibaba Group, has released Qwen2, the latest iteration of its open-source Tongyi Qianwen large language model (LLM) family. The AI model boasts impressive capabilities, challenging Meta’s Llama-3 in benchmarks.
The Qwen2 model comes in five variations, ranging from 0.5 billion to 72 billion parameters. Remarkably, it has been pre-trained on data covering 27 languages in addition to Chinese and English. Qwen2’s multilingual proficiency allows it to excel in tasks, including mathematics, programming, natural and social sciences, engineering, and humanities.
While Alibaba claims the high-end Qwen2-72B model consistently outperformed Meta’s Llama 3-70B in their internal benchmark tests, the independent Elo Arena benchmark ranks Qwen2-72B-Instruct below Llama3 70B. However, Qwen2-72B-Instruct is challenging Llama3 70B as the second-highest-ranked open-source LLM currently available based on evaluations by human testers.
These benchmark tests from Alibaba include nearly all aspects of functionality, establishing Qwen2 as a formidable challenge to other open-source competitors.
Qwen2 has the advantage of a wide context window, enabling it to operate with up to 128K tokens of context. This capability brings it to par with OpenAI’s GPT-4o and means the tool is well-suited to address tasks where long-form content needs to be processed.
In addition, Qwen2 achieved impressive results in the ‘Needle in a Haystack’ test that evaluates a model’s ability to extract relevant information from a large context without errors. Alibaba claims that Qwen2-72B-Instruct passed this test almost perfectly, further showcasing its prowess.
However, Alibaba has adopted the Apache 2.0 license for most Qwen2 models, aligning with standard open-source software practices.