Physicists' Old Trick for Magnets Could Make AI Models Half the Size Without Losing Much Brain

Key points

• Multiverse Computing's new method, published on Hugging Face, cuts Llama-3.3-70B (a large, publicly available AI language model) roughly in half and still scores nearly 23 percentage points higher on a standard knowledge test than the best rival compression technique.
• The approach borrows a physics equation called an Ising glass, normally used to model how atoms behave in disordered magnets, and uses it to score billions of possible ways to shrink an AI model.
• Existing compression tools typically score each section of a model in isolation; this method accounts for how sections interact, which turns out to matter a great deal at heavy compression.
• A model that runs on half the hardware draws roughly half the power, making this directly relevant to the data-centre electricity crunch AI2Day has been tracking.

Running a large AI language model, the kind of software that powers chatbots and coding assistants, is expensive. It demands specialist chips, large amounts of memory, and a great deal of electricity. One of the most direct ways to bring those costs down is to delete sections of the model entirely.

But deleting the wrong sections breaks the model. Whether any one section is safe to delete depends on what else you remove alongside it, making the search space grow exponentially. A 70-layer model has more possible combinations of sections to remove than there are atoms in the observable universe.

How does physics help with an AI problem?

Multiverse Computing realised this is exactly the kind of problem that physicists have been solving for decades with a model called an Ising glass. An Ising glass describes a disordered magnet where each atom can point up or down and every atom's behaviour is influenced by every other. Finding the lowest-energy arrangement of those atoms is the same kind of search as finding the best set of AI sections to delete.

The team assigned each section of the model a binary value, keep or delete, mirroring the up-or-down spins of atoms in the magnet. They then computed a matrix of couplings, a table describing how much each pair of sections influences each other, using a single pass through a small sample of text. That computation happens once and can be reused for any target size.

With the couplings in place, evaluating any candidate deletion set becomes a single cheap arithmetic step rather than a full model run. The team checked tens of billions of combinations on a single graphics card. For the largest case they tackled exhaustively, removing 8 of Llama-3.3-70B's 80 sections, that search took roughly two days. For cases too large to check exhaustively, they handed the problem to open-source optimisation solvers, including tools built for quantum annealing (a specialised search technique inspired by quantum physics). A solver called tabu search reached good answers in seconds.

What did the deleted sections actually look like?

The results challenged a common assumption in model compression. Most practitioners expect the best deletions to form a single continuous block of middle or late sections. The Multiverse method found that one of its strongest configurations removed a section near the very beginning of the model, something no existing heuristic would have proposed.

After a short retraining pass on that configuration, it outperformed even the lowest-energy solution on several benchmarks. The best answer wasn't the "ground state" the solver found first; it was the 17th best candidate. The value, then, is in generating a shortlist of good options rather than hunting for a single perfect answer, which is a much easier task for any solver.

At 50 percent compression of Llama-3.3-70B, the method scores close to 23 percentage points higher on MMLU (a standard multiple-choice test covering science, law, history and medicine) than the best competing block-removal technique.

Smaller, cheaper-to-run models are one direct answer to the data-centre power crunch. We covered two sides of that pressure earlier this month: the demand surge threatening to outpace grid capacity and projections that US data centres could consume more natural gas than Germany and Japan combined by 2035. A model running on half the hardware needs no new energy source.

Multiverse is a quantum-computing company applying its solver expertise to a classical AI problem, and the result is competitive with methods developed by teams far larger. If the approach generalises cleanly to newer model families, it could quietly reshape how companies decide what hardware they actually need. That's the thing to watch.

Common questions

Does this mean AI models will get worse to save power?

Not necessarily. At 50 percent compression the method loses some accuracy, but far less than rival techniques. Lighter compression, say 20 to 30 percent, typically costs very little quality while cutting memory and running costs noticeably.

Can this be applied to any AI model?

The paper demonstrates it on publicly available models in the Llama family. Each new model needs its own one-time coupling computation, but the underlying maths applies to any transformer-based model, the architecture used by most modern language AI.