Leech Lattice Vector Quantization: Using 24-Dimensional Geometry to Compress LLMs

*(English translation of a Chinese-language explainer post from zhichai.net.)*

Introduction

Large language models contain billions—even hundreds of billions—of parameters, making storage and compute costs enormous. In March 2026, a paper from Qualcomm AI Research (Tycho van der Ouderaa and colleagues) applied one of mathematics' most beautiful structures—the Leech lattice, the optimal sphere packing in 24-dimensional space—to LLM quantization.

Background: The Sphere Packing Problem

• 1-D/2-D: packing is trivial or hexagonal (~90.69% density, the honeycomb pattern).
• 3-D: Kepler conjectured the face-centered cubic packing (~74.05%) in 1611; proven by Thomas Hales in 1998.
• 8-D and 24-D: In 2017, Maryna Viazovska proved E8 is optimal in 8 dimensions; with Henry Cohn and others, she proved the same for the Leech lattice in 24 dimensions. She won the 2022 Fields Medal for this work.

The Leech lattice (John Leech, 1965) has remarkable properties:

1. Optimal packing in 24 dimensions. 2. Kissing number 196,560 — each sphere touches nearly 200,000 neighbors, ~10× random packing. 3. No short vectors (after normalization, no points within distance 2 of the origin). 4. Extreme symmetry via the Conway group Co₀ (over 8×10¹⁸ elements), linked to sporadic simple groups. 5. Universal optimality (Cohn et al., 2019): optimal for a broad class of energy-minimization problems.

It can be constructed from the extended binary Golay code, a perfect error-correcting code that corrects 3 errors per 24-bit block.

The LLM Compression Problem

A Llama 3.1 405B model stored in FP16 requires ~810GB. Quantization reduces parameters from 16-bit floats to 4-, 3-, or 2-bit representations at the cost of accuracy.

• Scalar quantization (each parameter independently) hits a fundamental information-theoretic limit.
• Vector quantization (VQ) groups parameters for joint compression but faces codebook explosion (2⁴⁸ entries for 24 parameters × 2 bits), intractable nearest-neighbor search, and memory bandwidth issues.

Prior work: QuIP# (2024) used Hadamard incoherence processing plus E8-lattice codebooks; QTIP used trellis-coded quantization (TCQ).

The LLVQ Method

LLVQ (Leech Lattice Vector Quantization) asks: use 24-dimensional Leech instead of 8-dimensional E8. Higher dimension yields better rate-distortion trade-offs, and Gaussian-like post-Hadamard weight distributions pair naturally with the lattice's geometry.

Key technical solutions:

• Extended Golay code search: exploits the Golay–Leech connection for efficient nearest-neighbor search without enumeration.
• Indexing and shell (angle) search: direct lattice-point-to-bitstring indexing and flexible rate allocation across shells.
• Fully parallelized dequantization kernel: a GPU kernel that converts bitstrings back to floating-point weights at inference speed comparable to QuIP#.

Results

Benchmarked against QuIP#, QTIP, AQLM, GPTVQ, and PVQ on WikiText-2/C4 perplexity and zero-shot tasks:

1. Lowest perplexity across Llama 2 and Llama 3 families at 2, 3, and 4 bits. 2. Largest gains at 2-bit extreme compression, where scalar methods essentially fail. 3. Biggest advantage on 405B-scale models. 4. Inference speed comparable to QuIP#.

Example (Llama 2 7B at 2 bits, WikiText-2 perplexity): FP16 ≈ 5.1; QuIP# ≈ 8.5; LLVQ ≈ 8.0 — better quality at the same 8× compression.

Outlook

• Higher-dimensional lattices (48-D, 72-D) with other extremal structures.
• Activation quantization alongside weight quantization.
• Quantization-aware training vs. post-training quantization.
• Hardware co-design for lattice operations.
• Extension to multimodal models.

References

1. van der Ouderaa, T. F. A., van Baalen, M., Whatmough, P., & Nagel, M. (2026). Leech Lattice Vector Quantization for Efficient LLM Compression. *arXiv:2603.11021*. https://arxiv.org/abs/2603.11021 2. Cohn, H., Kumar, A., Miller, S. D., Radchenko, D., & Viazovska, M. (2017). The sphere packing problem in dimension 24. *Annals of Mathematics*, 185(3), 1017-1033. https://doi.org/10.4007/annals.2017.185.3.8 3. Tseng, A., Chee, J., Sun, Q., Kuleshov, V., & De Sa, C. (2024). QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks. *ICML 2024*. https://arxiv.org/abs/2402.04396 4. Chee, J., Cai, Y., Kuleshov, V., & De Sa, C. (2024). QTIP: Quantization with Trellises and Incoherence Processing. *NeurIPS 2024*. https://arxiv.org/abs/2406.11235 5. Conway, J. H., & Sloane, N. J. A. (2013). *Sphere Packings, Lattices and Groups* (3rd ed.). Springer. ISBN: 978-1-4757-6568-7