Alibaba Cloud Unveils Qwen-7B Models to Rival Meta's Llama 2 AI

Alibaba Cloud is setting out to position itself as a formidable contender in the world of artificial intelligence, by launching two cutting-edge open-sourced AI models: Qwen-7B and its conversationally fine-tuned counterpart, Qwen-7B-Chat. This strategic move seeks to challenge Meta's recently open-sourced Llama 2 model.

Alibaba Cloud's emphasis on democratizing AI technologies is evident. Both models, along with their code, model weights, and documentation, are freely available to academics, researchers, and commercial entities around the globe. Companies with under 100 million monthly active users can utilize these models at no cost, even for commercial purposes. For larger programs, Alibaba Cloud offers a licensing option.

The Qwen-7B model is a culmination of extensive training on over two trillion tokens, encompassing languages like Chinese and English, and spanning various domains such as professional fields, coding, and mathematics.

The context length of the model reaches an impressive 8K. The Qwen-7B-Chat variant aligns closely with human instructions, making it adept at interactive tasks.

What sets the Qwen-7B model apart is its performance in the Massive Multi-task Language Understanding (MMLU) benchmark, scoring a remarkable 56.7 out of 100. This score surpasses several other major pre-trained open-source models, irrespective of their sizes. In the realm of Chinese foundational models, Qwen-7B also stands tall in the C-Eval leaderboard.

Meanwhile, Meta's Llama 2, released last month, shows an increase in 40% more training data compared to its predecessor, Llama 1, and exhibits commendable performance in various benchmarks. The introduction of these models follows Alibaba Cloud's earlier release of its proprietary LLM, Tongyi Qianwen, in April.