How to build AI more like software
IBM Research
IBM built a way to snap AI models together like software parts instead of one giant blob. Small models can now beat giants on narrow tasks, for way less money.
Based on reporting by IBM Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For years, working with a large language model has felt less like engineering and more like negotiation. You'd write a longer prompt, cross your fingers, and hope the model behaved. IBM Research thinks that's backwards, and its new release tries to prove it by treating models less like a slab of clay and more like a pile of LEGO bricks.
The centerpiece is something called Granite Libraries, a set of small, purpose-built adapters that snap onto IBM's Granite 4.1 models to handle one job really well — scoring a document for relevance, catching a hallucination, rewriting a query, making a safety call. IBM's Luis Lastras put it bluntly: models are just code with a lot more data than code, and the industry never bothered applying software's own lessons to them. So instead of retraining an entire model or drowning it in instructions, developers can bolt on an adapter function with a defined input and output, the same way you'd call a library in Python.
The numbers back up the pitch. Granite 4.1's 3B model, prompted the old-fashioned way to check whether a response meets certain requirements, scores 51% balanced accuracy on IFEval. Slap on the new requirement-check adapter, and that jumps to 84%, without touching the base model's weights. That's a small model suddenly punching in a different weight class, at a fraction of the inference cost of a giant generalist system.
Making all these adapters play nicely together is the job of Project Granite Switch, an experimental toolkit that inserts a new switching layer into the base model so it can flip between specialized experts on the fly. Pair that with activated LoRA, which lets the model carry its short-term memory from one step to the next instead of recalculating everything from scratch, and multi-step pipelines — safety check, retrieval, verification — stop grinding to a halt every time the model swaps hats. IBM's open-source Mellea library glues it all together, quietly inserting the right tags and enforcing formatting so the rest of the application never has to deal with raw, unpredictable model text.
None of this happens in a vacuum. Granite 4.1 itself is IBM's strongest release yet, with an 8B model matching its old 32B mixture-of-experts system and a 30B model going toe-to-toe with Llama 3.3 70B on enterprise benchmarks, all trained on roughly 15 trillion tokens and shipped under Apache 2.0. IBM isn't claiming modularity fixes everything about deploying generative AI at scale. But it's a concrete bet that the next real gains in enterprise AI come less from bigger models and more from better plumbing.
My take — AI-written commentary, not fact-checked reporting
This is the most sensible thing I've read out of a major AI lab in months, mostly because it isn't chasing a bigger frontier model — it's admitting that raw scale was never going to solve enterprise reliability problems. Open-sourcing the adapters and libraries under Apache 2.0 is the right call too; if generative computing actually takes off, it should be because developers everywhere can poke at it, not because IBM gatekeeps it behind an API. My only worry is that 'modular AI' becomes the next buzzword teams bolt onto a slide deck without doing the unglamorous work of actually training good adapters — the tech here is real, the discipline to use it well is not guaranteed.
Read more about this at: IBM Research
