Finance Association, May 20 (Editor: Zhou Ziyi) Many technology companies around the world have spent tens of billions of dollars to build various large-scale language models to power generative artificial intelligence products, and now, these companies are beginning to hope for a new way to drive revenue growth — small language models.

Small language models have far fewer parameters than large language models, but they still have powerful functionality. Microsoft, Meta, and Google have all recently released new small-parameter models for artificial intelligence.

Generally speaking, the greater the number of parameters, the better the performance of artificial intelligence software, and the more complex and ingenious the tasks it can perform. Last week, OpenAI's newly announced latest model GPT-4O and Google's Gemini 1.5 Pro are estimated to have more than 1 trillion parameters, while Meta's open source Llama model has about 400 billion parameters.

However, the computational power required to run large language models is enormous, which means that it is expensive. In addition to being difficult to persuade some enterprise customers to pay large operating expenses, data and copyright issues have also hindered the use of artificial intelligence products.

alternatives

Some tech companies are currently promoting small language models with only a few billion parameters as cheaper, more energy efficient, and customizable alternatives. These models can be trained and run with less power, and they also protect sensitive data.

Google, Meta, Microsoft, and French startup Mistral have successively released their own small language models, which show advanced functionality and can better focus on specific application features.

Nick Clegg, president of global affairs at Meta, said bluntly that Llama 3's latest 8 billion parameter model is comparable to GPT-4. He notes, “I think you've seen excellent performance in almost every measure you can think of.”

Microsoft, on the other hand, said that the performance of its small Phi-3 model with 7 billion parameters is superior to GPT-3.5, an earlier version of the OpenAI model.

Eric Boyd, vice president of Microsoft's Azure artificial intelligence platform, said, “By getting this high quality at a lower cost, you're actually providing your customers with more apps that allow them to do things that are really unbelievable.”

Boyd also notes that “smaller models will bring interesting applications and can be extended to mobile phones and laptops.”

Another major advantage of the smaller model is that tasks can be processed “locally” on the device rather than sending information to the cloud, which may appeal to customers who care about information privacy.

Currently, Google's “Gemini Nano” model is embedded in Google's latest Pixel phone and Samsung's latest S24 smartphone; Apple also hinted that it is also developing an AI model to run on the iPhone. Last month, Apple released the openELM model, a small model designed to perform text-based tasks.