The low-code artificial intelligence development platform company Predibase has announced that it will launch a collection of at least 25 open-source and fine-tuned large language models (LLMs) that it claims can rival or even outperform OpenAI's GPT-4.

The LLM collection, called LoRA Land, aims to meet the needs of use cases such as text summarization and code generation. Predibase claims to provide a more cost-effective way for enterprises to train high-precision, specialized generative AI applications.

The company raised $12.2 million in an expanded Series A funding round in May last year. As the creator of a low-code machine learning development platform, Predibase enables developers to build, iterate, and deploy powerful AI models and applications at lower costs. The startup aims to help small companies compete with major AI companies like OpenAI and Google LLC by replacing complex machine learning tools with an easy-to-use framework.

Using Predibase's platform, teams only need to define the content they want the AI model to predict using pre-built LLMs, and the platform takes care of the rest. Novice users can choose entry-level models from a variety of recommended architectures, while experienced practitioners can fine-tune the parameters of any AI model using the tools provided. Predibase claims that its tools can launch and run AI applications from scratch in just a few days.

Predibase states that with the launch of LoRA Land, the company will be able to economically and efficiently provide multiple fine-tuned LLMs on a single graphics processing unit (GPU). LoRA Land LLMs are built on the open-source LoRAX framework and Predibase's serverless fine-tuning endpoints, each tailored to specific use cases.

Building GPT models from scratch or fine-tuning existing LLMs with billions of parameters can be extremely costly. Therefore, smaller and more specialized LLMs have become a popular alternative, allowing developers to create high-performance AI applications at low costs using parameter-efficient fine-tuning and low-rank adaptation methods. Predibase has incorporated these techniques into its fine-tuning platform, allowing customers to select the most suitable LLM for their use case and fine-tune it in an affordable manner.

To prove its point, Predibase claims that the average GPU cost for the 25 LLMs in LoRA Land is less than $8. This means that customers will be able to fine-tune hundreds of potential LLMs on a single GPU using LoRA Land, which is not only cheaper but also allows for faster testing and iteration as there is no need to wait for a cold GPU to start before fine-tuning each model.

Andy Thurai, Vice President and Chief Analyst at Constellation Research Inc., believes that Predibase offers a compelling product considering the typically high costs of implementing AI. He explains that while the initial experimentation cost of accessing LLMs through APIs is relatively low, costs quickly escalate when deploying comprehensive AI implementations.

Thurai adds, "From a resource perspective, another alternative involving fine-tuning open-source LLMs can also be quite expensive and challenging in terms of skills, which poses a problem for companies without qualified AI engineers." He believes that Predibase now offers a third option with a set of 25 fine-tuned LLMs that can be further refined and deployed on a single GPU.

Thurai sees this as an interesting idea that could have a significant impact on small companies, as many small, specialized models have already demonstrated their ability to surpass large LLMs in certain specific use cases. "The desire to use open-source LLMs and the limited availability of AI skills could have a significant impact on companies considering this from that perspective," Thurai says. "If companies decide to use different fine-tuned models for each use case, Predibase's offering could be highly sought after."

The company's serverless fine-tuning endpoint deployment option means that customers can even create AI models without the need for GPU resources, significantly reducing operating costs, according to the analyst. "While it remains to be seen whether Predibase's models perform better than GPT-4, it sounds like a very appealing alternative for many AI applications," Thurai says.

Dev Rishi, Co-founder and CEO of Predibase, states that some of the company's customers have already recognized the benefits of using smaller, fine-tuned LLMs for different applications. One of these customers is AI startup Enric.ai Inc., which provides a platform for coaches and educators to create AI chatbots that incorporate text, images, and voice.

"This requires the use of LLMs for many use cases, such as translation, intent classification, and generation," says Andres Restrepo, CEO of Enric.ai. "By switching from OpenAI to Predibase, we were able to fine-tune and serve many professional open-source models in real-time, saving over $1 million annually while creating engaging experiences for our audience. Most importantly, we have ownership of these models."

Developers can start fine-tuning LoRA Land LLMs today with Predibase's free trial product. The company also offers free developer tiers (with limited resources) and paid options for larger projects.