• Multiverse Computing's Ising-based block-removal method beats a block-influence baseline by nearly 23 MMLU points (76.9 vs. 54.0) on Llama-3.3-70B-Instruct at 50% depth compression, per its own arXiv preprint 2602.00161.
• The pruned 70B model still trails the dense original by more than five MMLU points (76.9 vs. 82.2), so the win is relative to a weak baseline, not proof of near-lossless compression.
• The advantage shrinks to merely comparable with prior leading methods on the smaller Llama-3.1-8B-Instruct, and a mathematically suboptimal configuration beat the 'optimal' one after retraining — undercutting the core low-energy-predicts-accuracy premise.
• No independent lab has reproduced any of the reported figures; every number traces back to Multiverse Computing, which closed a $570 million Series C at a $1.7 billion valuation around the same period.
• Brute-force validation is computationally prohibitive — roughly 29 billion configurations and about two GPU-days just to check 8-of-80 block removals — which is why the company substitutes a tabu-search solver instead of exhaustive search.

Multiverse Computing, the Spanish quantum-inspired compression startup behind CompactifAI, published a Hugging Face blog post and companion arXiv preprint (2602.00161, last revised September 11, 2026) proposing a new way to pick which transformer blocks to delete when shrinking a large language model. Instead of scoring blocks one at a time, the method treats block selection as a constrained binary optimization problem mapped onto an Ising glass — the classic statistical-physics model of interacting spins — and reports that the resulting pruned Llama-3.3-70B-Instruct beats the strongest prior block-removal baseline by nearly 23 MMLU points at 50% depth compression. Every one of those numbers comes from Multiverse Computing itself; no outside lab has reproduced them yet.

How the Ising framing actually works

The method runs one calibration pass to build a Hessian matrix over all transformer blocks: diagonal entries score each block's individual importance, off-diagonal entries capture how removing two blocks together compounds the damage. That coupling matrix becomes the interaction terms of an Ising glass, and minimizing its energy is treated as a proxy for downstream benchmark accuracy. Brute-force verification is intractable — checking just 8-of-80 block removals took roughly two GPU-days and about 29 billion configurations — so Multiverse instead runs an open-source tabu-search solver that finds low-energy configurations in seconds.

The headline number, and what it leaves out

At 40-of-80 blocks removed (50% depth) with no retraining, Multiverse reports Llama-3.3-70B-Instruct scoring 76.9 MMLU under its Ising-optimized block set versus 54.0 for the block-influence baseline — the source of the 'almost 23 percentage points' claim. That comparison is real but incomplete: the dense, unpruned model scores 82.2 MMLU, so even the winning configuration gives up more than five points of accuracy. The baseline it beats is also known to be weak specifically at deep, no-retrain compression ratios — the exact regime where a smarter search should look best.

Where the physics metaphor stops holding

The advantage narrows sharply at smaller scale: on Llama-3.1-8B-Instruct, Multiverse's own results describe the method as performing merely comparably with prior leading pruning approaches, not beating them. More telling, the paper reports that after light retraining on the 8B model, a deliberately suboptimal configuration — the 17th-lowest-energy state, not the ground state — outperformed the mathematically 'best' block set. That result undercuts the core premise that low Ising energy predicts post-training accuracy, which matters because real deployments apply retraining or distillation 'healing' after pruning.

Who should use it, who should skip it

Teams doing single-shot, no-retrain compression of very large (70B-plus) dense models, with in-house capacity to compute a full pairwise-block Hessian and run the open-sourced solver, get a documented method that beats a known-weak baseline by a wide margin. Everyone else — teams pruning 8B-and-under models, or anyone who retrains after pruning, which covers nearly every production compression pipeline — should not expect the headline 23-point gain to show up in their own numbers, per Multiverse's own smaller-model and post-retraining results.

The bottom line

Bottom line: Wait. The Ising-CBO reformulation is a genuine methodological contribution — pairwise block coupling via a Hessian, plus a solver that beats naive greedy block-importance ranking, is real progress on a real combinatorial problem, and the code is open. But every number in this post comes from Multiverse Computing itself, announced around the time of a $570 million Series C at a $1.7 billion valuation, with no independent lab having reproduced the 70B result. Until someone outside the company reruns the 40-of-80 Llama-3.3-70B comparison and confirms 76.9-versus-54.0 MMLU, treat this as a promising research signal, not a production pruning recipe.