I’ve spent the past few years building machine learning models, and if there’s one thing that keeps driving me insane, it’s this: people throwing LLMs at every single problem like it’s the only tool in the toolbox.
You need to classify documents? LLM.
You need to predict customer churn? LLM.
You need to detect fraud in structured transaction data? LLM.
Look, I get it—LLMs are cool. But they’re also expensive, slow, and often wildly inefficient for most use cases. The fact that a model trained on all of human knowledge is being used to determine whether an email is spam just feels… wasteful.
Most real-world AI problems don’t need a 100B parameter behemoth—they need small, efficient, and specialized models that actually fit the task.
So a friend and I decided to stop complaining and build something—a tool that actually helps people build task-specific models without needing ML expertise or massive datasets. It’s called smolmodels, and it’s open-source. Instead of throwing GPT at your problem, you just describe what you need, and it builds a model for you.
I honestly think the future of AI isn’t in making bigger models, but in making ML more accessible and practical for real-world tasks. Not everything needs to be a transformer with trillion-dollar compute bills attached.
Not Every AI Problem Needs a 100B Parameter Model 🤦‍♂️
byu/Pale-Show-2469 inFuturology
12 Comments
Small models are OP when you need to do domain specific tasks. They are also radically cheaper, and if you want to make a product that’s sustainable, you’ll beat out a competitor who’s blindly using an openai API or hosting llama or some OSS large model.
I don’t know if it’s just me, but you doc seems to be linking to the same welcome page, whatever navigation I try to take.
Very interesting! For Plexe, it says “get started for free”, but what are the different potential costs later on?
This is just so much blabla.
With the speed of development in technology, this might solve a problem for the next 3 years.
I read that whole comment wondering why anyone would spend 100B on a limited liability company to protect AI software that may under perform. Then I got to the end and realized the M stands for Model…
I just had similar thoughts. While AI has definitely given me any new opportunities (automate stuff with python with 0 coding background) I feel like people are trying too hard to make agents / use LLM everywhere while there are just simpler and more elegant solutions for their minor problems (which are still solvable with AI but just a different approach)
Does the doctor anayzing your blood tests have a team of physicists, meteorologists, and software engineers helping with that? No? Then why should AI do that? I think you’re right that small models are the future. In fact I think large models are a function of the technology not being fully matured, and the big labs will move away from them in time. It makes more sense to have a ton of small expert models that all collaborate, dynamically loading and unloaded as needed. AGI at home.
I work with politicians the amount of time I have to remind them; just because you have a hammer doesn’t make the problem a nail is tiering…
Supports self-hosted llms like ollama or only 3rd party providers?
B-but it is impossible to make a linear regression model with less parameters…
This is very interesting.
I was hoping to know – what is the model under the hood? How many parameters do you actually have?
Or to rephrase, is the resultant model (your app builds) a vector DB, or like a 80M param LLM (or something).
I took a look at the docs but I couldn’t discover the answer.
—
If it helps to know my reasoning; This seems great to produce micro models for self-service. In a [cluster](https://github.com/Strangemother/clam/tree/main/v3) of trained small models with mini tasks.
I feel you on the LLM overkill! As someone who’s dabbled in ML, it’s refreshing to see practical solutions like smolmodels. It reminds me of how I’ve been using BeFreed.ai for personalized book recommendations and summaries. Instead of throwing a massive language model at everything, it uses targeted AI to suggest books based on my goals and interests. It’s like having a smart reading buddy that knows exactly what I need. Maybe the future of AI is indeed about making things more accessible and tailored, whether it’s building models or enhancing our learning experiences.