A World of Small AI Models
The inevitable shift from large, foundational models to tiny, task-specific ones.
Small AI models are underrated, yet immensely capable. While most AI companies are focused on training large “frontier” models capable of general reasoning, we think there is still a great deal of potential to be unlocked by training smaller models that are quick to run and highly accurate at narrow tasks. Small models can also be run locally, on a computer or smartphone, and thus maintain a user’s privacy.
We’re building MiniBase.ai to help anyone design, train, and deploy small models.
Such “tiny” models have been used for decades. Your phone’s keyboard is running a small language model to guess the next word as you type. Noise-cancelling headphones run a model that predicts and subtracts background noise in real time. Cameras in phones use small vision models to detect faces and adjust focus. Email services run spam filters trained to recognize patterns. And so on.
Small models — often under 10 GB in size — can also outperform large models on narrow tasks. The Orca 2 model (~5 GB) outperforms models 5–10× larger on complex reasoning benchmarks in zero‑shot scenarios. The Phi‑3 family of models (as small as ~1.8 GB) consistently beat slightly larger models in benchmarks covering language, coding, and math.
Even the frontier laboratories are releasing smaller, open-weight models. OpenAI recently released their gpt-oss models, for example, under a permissive Apache license. The smaller version only has 21 billion parameters (~16 GB) but performs comparably to o3-mini on most tasks.
As more and more developers turn to small models, it’s inevitable that more companies will pop up to serve them, too. HuggingFace already offers more than 600,000 models, many of which are small enough to run on a laptop. Their community has also uploaded thousands of datasets and training scripts to reproduce those models. Databricks, similarly, has tools to fine-tune and deploy models. Both of these companies make it possible for a small team to go from raw data to a deployed model without owning a GPU cluster.
But although Hugging Face and Databricks are extremely powerful, they are marketed primarily at developers. If you were, say, a student who wanted to train a model to identify bird species, run it on a Raspberry Pi, hook everything up to a camera, and then mount it on a bird house — where would you start?
The model names on Hugging Face are esoteric and difficult to parse. (What’s the difference between Phi-3 and Phi-3.5-mini-instruct? ) It’s often unclear what dataset is best-suited to training, or how that dataset should even be structured or collected in the first place. And even if these two problems are resolved, you still need to deal with Python environments, GPU drivers, and memory limits. There is a lot of troubleshooting required.
MiniBase.ai is designed to solve these problems. We’re building the tools and infrastructure required to make data cleaning and model training as painless as possible. With our tools, absolutely anyone can train, fine-tune, and deploy their own tiny models. We are also growing a community around these tools, where users can be paid for their datasets and models.
We’ve spent the last few months talking to dozens of hobbyist model builders. We’ve helped scientists build “fraud detector” tools that speed up their searches for papers with statistical or image anomalies. And we’ve talked to founders at companies that have already trained models in-house, but now just want an easy solution to re-train or fine-tune those older models. Over time, we’ve come to believe that there’s a huge gap in the market for a community-based effort focused entirely on small AI models. So come join our community on Discord!
In the future, small models will be everywhere, running on whatever device happens to be nearby. Farmers will run small models on solar-powered sensors to track soil conditions without an internet connection. Office workers will carry models on USB sticks to automate dull parts of their job. These models won’t be remarkable for what they are, but for how common they’ve become; quiet pieces of software that do one job well.
— The MiniBase team