Chinese tech giant Alibaba is bringing Meta’s new open-source large language model to its cloud computing platform. The Llama 2 model has just been made available free of charge for both commercial and academic use, and is already running on other cloud platforms, as well as being available to download and run locally.
Meta published the code and weights for its model last week, partnering with Microsoft on the initial release. While Microsoft doesn’t publish its own models directly, it is heavily invested in OpenAI and uses the AI lab’s models in its own products. But unlike OpenAI, Meta is making all its trained models available open source.
Alibaba says the entire Llama 2 series of large language models would be available in China for the first time, including for training and deployment through Alibaba Cloud. It will allow companies in China to build new tools using the AI model as a base. Other models from OpenAI and Google are not widely available in the country.
Organisations deploying the Meta model will have to obtain a licence from the social media giant if they have more than 700 million active monthly users. They will also have to comply with Chinese AI regulation rules, including ensuring any output meets certain standards and doesn’t go against local censorship laws.
Being open source allows Chinese companies to fine-tune, train and customise the AI model more than is possible with models from other vendors. This could help them put in place stricter guardrails that meet Chinese censorship requirements.
Alibaba AI: systems must meet ‘core values of socialism’
China’s ground rules for the use of AI include ensuring output “aligns with the core values of socialism” and doesn’t subvert state power. They also place a burden on developers to ensure data doesn’t discriminate against ethnicity, race or gender or provide false information. Moderation rules in the guidelines place a requirement on providers to ensure content is consistent with “social order and societal morals”, doesn’t endanger national security, avoids discrimination, is accurate and “respects intellectual property”.
Launching on Alibaba Cloud puts Meta’s AI model directly up against dozens of homegrown Chinese AI models, including those built by Alibaba rivals Tencent and Baidu. A report by the Chinese Ministry of Science and Technology published in May found that Chinese organisations had launched 79 large language models since 2020 as they try to catch up with Western counterparts.
Llama 2 is likely to be widely deployed due to the ease of fine-tuning and customising the model. It can be run on-premises or through cloud platforms and comes with code, model weights, pre-trained data dn a fine-tuned version for conversations like those used for ChatGPT, Bard and Claude.
This ability to run more widely has also led Qualcomm to confirm it would be compatible with its Snapdragon chips, opening up the possibility of models running natively on Android and Windows devices.
“We applaud Meta’s approach to open and responsible AI and are committed to driving innovation and reducing barriers-to-entry for developers of any size by bringing generative AI on-device,” said Durga Malladi, the chipmaker’s senior vice-president and general manager of technology, planning and edge solutions businesses.