3 tips for navigating the open-source AI swarm - 4M models and counting

KTSDESIGN/SCIENCE PHOTO LIBRARY/Science Photo Library via Getty Follow ZDNET: Add us as a preferred source on Google. ZDNET's key takeaways Open source dominates software technologies; AI is no exception. Hugging Face now has 4 million AI models in its library. Small language models and open-source AI hardware are emerging. You could almost call it too much of a good thing. AI models have gone open source -- en masse -- to the tune of millions now competing for a place in our applications. There are echoes of the open sourcing of vital enterprise software and tools, with the opening of operating systems, databases, development platforms, and utilities at no cost. There is a difference this time, however. The AI world has virtually exploded into a mind-blowing array of models that handle everything from neural networks to risk-management tools. Open-source AI models and software enable users to download and experiment, at little to no cost, with the right models to serve their business requirements. And as these requirements change, businesses can quickly adopt alternative solutions with little fear of lock-in. The grand repository for open-source AI models is Hugging Face, which now maintains four million open-source AI models, built by and serving a community of 10 million AI developers. Also: Google's new Jules Tools is very cool - how I'm using it and other Gemini AI CLIs "There is one new model being published every five seconds," Thomas Wolf, cofounder and chief science officer of Hugging Face, pointed out in a recent interview with MIT's "Me, Myself, and AI" series hosts. Notable Hugging Face offerings include the Meta series, the Llama series, and DeepSeek. The challenge is sifting through what has become a bewildering array of models to identify and try out the ones that fit your business or even personal projects. And, as was and is the case with earlier open-source software, working with these models is an "as-is" proposition, and not for the technically inexperienced or faint of heart. Also: Despite AI-related job loss fears, tech hiring holds steady - and here are the most in-demand skills The following are five takeaways for working and succeeding with today's nonstop gusher of open-source AI models: For a quick solution, consider commercial offerings Open-source AI is a long-term proposition. Closed or commercially available AI models attract funding and resources to get solutions out to market quickly -- but this only serves technologies in the early stages. Going back a couple of decades, vendors such as Microsoft and Apple offered support and hand-holding with their commercial operating systems, while open-source OSes such as Linux required interactions with community members. In its first decade or two, "Linux was somehow more for fanatics or geeks," said Wolf. "Now, Linux is really the basics of all enterprise software and all enterprise cloud." The same will evolve in the open-source AI space, he said. Want more stories about AI? Sign up for Innovation, our weekly newsletter. Don't underestimate small models You're likely not going to need gigantic large language models for narrow roles -- you can download fairly robust AI models that can run on laptops -- and even on mobile phones. Small language models are gaining traction as a more efficient means of performing specific tasks that don't require all the resources of LLMs. Hugging Face just published a new small language model, SmolLM3, with three billion parameters, that will run nicely on your personal devices, said Wolf. Get recommendations from your network Look to social media to help identify the right open-source AI model. As with most things in this world, we depend and trust the recommendations of trusted advisors -- be it friends, colleagues, influencers, or online community members -- to help us make decisions about the right product or solution. This certainly is the leading factor in decisions about mobile app downloads. The same would help with plowing through the four million options found at Hugging Face. The company initially attempted to curate its huge library of AI models through guidelines and manuals. "But with one new model every five seconds, that's not really sustainable," Wolf said. Filtering through searches is one approach to narrow down the field, but social media recommendations, blog posts, and adoption trend data can be the most productive way to winnow down choices. Want to follow my work? Add ZDNET as a trusted source on Google.