Alibaba Cloud has launched Qwen2, the latest iteration of its Tongyi Qianwen large language model (LLM) family. Qwen2 boasts a range of models from 0.5 billion to 72 billion parameters and has been trained on a diverse dataset covering 27 languages, including Chinese and English. It excels in various fields such as mathematics, programming, natural sciences, social sciences, engineering, and humanities.
Alibaba claims that their top model, Qwen2-72B, outperformed Meta’s Llama-3 in internal benchmark tests, although independent benchmarks like Elo Arena rank Qwen2-72B-Instruct slightly below Llama-3 70B. Nevertheless, Qwen2-72B-Instruct is competitive as the second-highest-ranked open-source LLM based on human evaluations.
Qwen2 supports a large context window of up to 128K tokens, similar to OpenAI’s GPT-4o, making it suitable for tasks requiring extensive processing of long-form content. It also performed exceptionally well in the ‘Needle in a Haystack’ test, demonstrating near-perfect accuracy in extracting relevant information from large contexts.
Despite its capabilities, Alibaba has adopted the Apache 2.0 license for most Qwen2 models, adhering to open-source software standards.
