A few months after releasing the large language model Tongyi Qianwen in April to compete against OpenAI's ChatGPT, Alibaba has revealed plans to make open source two of its models: a 7 billion-parameter model called Qwen-7B and a conversational app version called Qwen-7B-Chat.
Qwen-7B-Chat is a fine-tuned version of the first model that can conduct human-like conversations. Both models are smaller versions of Tongyi Qianwen.
These Alibaba generative AI models are designed to rival Facebook parent company Meta, which made its open source LLM, Llama 2, available for commercial use last month.
Along with Alibaba, Meta is one of the major technology vendors to opt for an open source LLM. Meanwhile Chat-GPT creator OpenAI and Google have chosen to not make the latest versions of their LLMs open source.
The two new models from Alibaba, released Aug. 4, are aimed at researchers and academics to train and are the smallest of Alibaba's family of LLMs, according to Constellation Research analyst Andy Thurai.
The vendor, the biggest cloud vendor in Asia, has no plans to make the larger models open source -- which differs from Meta's strategy of making all its LLMs open source. While it competes with Meta, Alibaba partners in making Meta's LLM available in Alibaba Cloud.
Despite its attempts to compete with U.S. vendors, Alibaba has not had much success, Thurai said.
Andy ThuraiAnalyst, Constellation Research
"They are most probably limited to their cloud users and mostly in China and some other Far East countries," he said, adding that the vendor is playing catch-up for the most part to other vendors in the market, such as Meta and Baidu, China's primary search engine vendor.
"While I see a market demand to produce Chinese text, they will be competing not only against Baidu but also against many other players in the market," Thurai continued.
However, Alibaba's LLMs enable customers to generate text in both Chinese and English, which is a notable differentiator, Thurai said.
Moreover, having its models means Alibaba does not have to use LLMs from vendors such as OpenAI in its products.
The models are also comparable to OpenAI's because customers can train or retrain the models using customers' proprietary data, similar to ChatGPT and to Baidu's ChatGPT equivalent, Ernie Bot.
While Qwen-7B and Qwen-7B-Chat are accessible to researchers and academics, organizations with more than 100 million monthly active users will still need a license from Alibaba before they can use the two models, according to the vendor.
Esther Ajao is a news writer covering artificial intelligence software and systems.