Alibaba open sourced the Tongyi Qwen2 model last week. It can be said to be the strongest open source model of this scale at the present stage. After its release, it directly ranked first on the Huggingface LLM open source model list, surpassing the newly released Llama3 and a host of other open source models.
Qwen2 is particularly outstanding in code and mathematics representing reasoning ability as well as long text performance. It has achieved excellent results in reasoning-related tests and needle-in-a-haystack tests.
Model Overview:
The Qwen 2 model consists of Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, Qwen2-57B-A14B, and Qwen2-72B. Among them, Qwen2-57B-A14B is a MoE model.
Based on Chinese and English corpora, high-quality data related to 27 languages has been added to the training data.
The supported context length has been increased, with a maximum of 128K tokens (Qwen2-72B-Instruct).
LLM Arena Results:
The results in the LLM Arena are also quite good, making it the best-performing open source model in Chinese currently (ranking 7th).
Compared with v1.5-110B, there is a significant improvement. In overall performance, it can be comparable to GPT-4-0314. In handling "difficult problems", it almost catches up with the best-performing open source model Llama-3-70B.
Model Download: https://huggingface.co/collections/Qwen/qwen2-6659360b33528ced941e557f