Block-pruning method halves Llama 3.3 70B with 23-point MMLU gain

single source· 1 articles · confidence: medium · first seen 2026-09-21 00:00 UTC

What this means for you

Nothing to act on yet. This is a blog post, not a release: no code, no weights, no evaluation date. The 23-point figure is the authors' own comparison against methods they selected, on one model at one compression ratio — read it if you prune large models, but do not change tooling on it.

Multiverse Computing has posted a paper compressing large language models by deleting whole transformer blocks rather than individual weights, and treating the choice of which blocks to remove as a constrained binary optimisation problem — the same mathematical shape as an Ising spin glass, a disordered magnet in which every spin interacts with every other. Modelling those interactions lets the authors rank candidate configurations without benchmarking each one. At 50% compression of Llama-3.3-70B-Instruct, the paper reports a gain of almost 23 percentage points on MMLU over the best competing block-removal method. The post gives no released code or weights, and no evaluation date for the comparison.

Models in this story

Key facts

  • ·The paper is titled "LLM Compression by Block Removal with Constrained Binary Optimization". source
  • ·At 50% compression of Llama-3.3-70B-Instruct, the authors report almost 23 percentage points gained on MMLU over the best competing block-removal method. source
  • ·The method assigns a binary variable to each transformer block and maps the selection problem onto an Ising spin glass with a fixed number of "up" spins. source
  • ·The energy of the spin system is used as a proxy for a pruned model's benchmark score, so candidate configurations can be ranked without benchmarking them. source
  • ·The post was published on the Hugging Face blog on 21 September 2026. source
  • ·The post states no evaluation date for the MMLU comparison and links no released code or weights. source

What the sources say

  • Hugging Face Blog — Authors set out their own block-removal compression method and the statistical-physics formulation behind it.

Sources

The original reporting. Follow these — they did the work.

← the wire