Alibaba says new AI model Qwen2 bests Meta’s Llama 3 in tasks like maths and coding

Alibaba said its latest large language model holds up against the most cutting-edge open-source models like those from Facebook owner Meta Platforms

Reading Time:2 minutes

Alibaba is touting the advantages of its latest Tongyi Qianwen model, saying it rivals the sophistication of Llama 3, the open-source model from Facebook owner Meta Platforms. Photo: Shutterstock

Published: 8:07pm, 7 Jun 2024Updated: 8:08pm, 7 Jun 2024

Alibaba Group Holding, the e-commerce giant that is investing heavily in generative artificial intelligence (AI), has updated its open-source AI models with claims of besting Meta Platforms’ Llama 3 model in certain tasks.

Alibaba Cloud on Friday launched Qwen2 – the second iteration of its open-source Tongyi Qianwen family of large language models (LLM), the technology behind chatbots such as OpenAI’s ChatGPT – with a series of updates including multilingual pre-training and an expanded the context window. That means it now allows for much longer queries and answers, putting in it the league of the world’s most powerful open source LLMs.

Qwen2 comes in five variations. The high-end Qwen2-72B model consistently offered better results than Meta’s Llama 3-70B – the strongest open-source AI model from the Facebook owner – in various benchmark tests, according to Alibaba. Tests include maths, coding, natural and social sciences, engineering and the humanities, the company said in a post published to the model’s official GitHub page.

Alibaba, which owns the South China Morning Post, has launched Qwen2 just one month after unveiling Tongyi Qianwen 2.5, which is closed source. The company said at the time that model performs better in various Chinese capabilities than GPT-4, OpenAI’s most advanced model, which is also closed source.

The five variations of Alibaba’s Qwen2, from the more nimble Qwen2-0.5B to its most sophisticated Qwen2-72B, have between 490 million and 72.7 billion parameters. They are also trained on 27 languages, in addition to Chinese and English: nine from Europe, four from the Middle East, and 14 from Asia.

The rapid follow-up launch of a new AI model, with capabilities matching leading global models, reflects the confidence that the Chinese firm has in funnelling an increasing amount of resources into an AI race that has engulfed much of the tech industry.