Apple researchers found a smarter way to build smaller AI models, and it could cut your phone's battery drain

Key points

• Apple ML Research published a paper describing IDEA Prune, a pipeline that pre-trains an oversized AI language model and then prunes it to a smaller, deployable size.
• Pruned models built this way outperform models trained at the target size from scratch, according to the paper's benchmarks.
• The technique targets devices with tight computing budgets, which could benefit smartphones and tablets where battery life and memory are limited.
• Building big first, even when that larger model is never shipped, pays off in the quality of the final small model.

Big AI models are impressive. They're also expensive to run, which is why researchers have spent years trying to squeeze the same quality into a smaller package. One approach, called structured pruning (cutting out whole sections of a model the way you'd trim fat from a piece of meat rather than picking out individual cells), has looked promising. The problem is that most pruning research starts with an already-trained model of the right size and prunes from there.

Apple ML Research took a different path: build something much bigger first, then cut it down.

Why build a model you never plan to ship?

It sounds wasteful. You train a large language model, the technology behind chatbots like ChatGPT and Claude, at a bigger size than you ever intend to deploy. Then you prune it.

The extra size during training acts like a scaffold. The model learns richer, more stable patterns when it has more room to work with, and those patterns survive in the smaller version after trimming. That produces better results than a model that was always the cramped size.

Think of it like learning to cook in a well-equipped professional kitchen before working in a tiny home galley. The skills carry over even when the space does not.

Apple ML Research has published steadily on on-device efficiency this year. We covered the lab's approach to teaching AI more languages cheaply on 20 August, and this paper sits in the same tradition: make a capable model small enough to live on your hardware.

What does this mean for everyday devices?

Running a capable AI model on a phone or a smart watch means making trade-offs right now: slower responses, hotter hardware, faster battery drain. IDEA Prune is designed for exactly that constraint, what researchers call a limited inference budget (a strict cap on how much computing power a device can spend generating a response).

If the technique holds up outside the lab, it could mean more capable on-device AI assistants that respond faster and use less power, without needing a constant connection to a remote server. That matters for privacy too: AI that runs entirely on your device doesn't send your questions or data to a cloud computer somewhere else.

What happens next?

This is a research result, not a product announcement. Apple hasn't said when or whether IDEA Prune will appear in shipping software.

Academic benchmarks, the standardised tests researchers use to compare models, don't always predict real-world performance. Models that shine in a paper sometimes disappoint in a product.

Still, the direction is clear. The industry is moving toward smaller, more efficient AI that runs close to the user, and techniques like this one are part of how that gets built.

One honest takeaway: if you buy a new phone in the next two years, the AI features on it will almost certainly owe something to this generation of pruning research. You don't need to understand the paper. You'll feel the results in battery life and response speed.