Alibaba Cloud's Qwen2-72B Tops Open Source Model List

TapTechNews on June 28, a tweet was posted by Clem Delangue, co-founder and CEO of HuggingFace, on platform X on June 26, saying that the open-source Tongyi Qianwen (Qwen) instruction fine-tuning model Qwen2-72B from Alibaba Cloud topped the list of open source models.

HuggingFace has announced a brand-new list of open-source large language models. By using 300 Nvidia H100 GPUs, it re-evaluated the current mainstream large language models such as MMLU-pro and other standard evaluations, and described Qwen2-72B as the king in its key introduction, and stated that many open-source models in China have a place on the list.

He stated that in order to provide a brand-new list of open-source large models, 300 H100s were used to conduct new evaluations for more than 100 mainstream open-source large models currently available globally, such as Qwen2, Llama-3, mixtral, Phi-3, etc. on benchmarks such as BBH, MUSR, MMLU-PRO, and GPQA.

The Qwen-272B model open-sourced by Alibaba stood out in the fierce competition, not only surpassing Meta's Llama-3, but also surpassing Mixtral from the well-known French large model platform Mistralai, becoming the new industry leader. TapTechNews quotes the official blog post and attaches the list ranking as follows:

Rank New List Ranking ⭐Qwen/Qwen2-72B-Instruct 2meta-llama/Meta-Llama-3-70B-Instruct 3 microsoft/Phi-3-medium-4k-instruct 401-ai/Yi-1.5-34B-Chat 5CohereForAI/c4ai-command-r-plus 6abacusai/Smaug-72B-v0.1 7 Qwen/Qwen1.5-110B 8 Qwen/Qwen1.5-110B-Chat 9 microsoft/Phi-3-small-128k-instruct 10 01-ai/Yi-1.5-9B-Chat