mlsanityData to 7 Oct 2026

Tied Trit-Planes: Constraining PTQTP to a Uniform Nine-Level Quantizer, with a Persistent Folded Format for Disk-Streamed Mixture-of-Experts Serving

arXiv 2608.08910 ↗First posted 9 Aug 2026cs.CL cs.LGcited by 0

On the map: Mixture-of-experts → MoE large language models

Abstract (arXiv)

PTQTP decomposes LLM weight matrices into two ternary (trit) planes with two free per-group scales. Tying the scales to a fixed ratio of three collapses the decomposition into a single uniform nine-level quantizer, a known balanced-ternary identity. To our knowledge, at the time of writing, this work is the first to impose that identity as a constraint inside PTQTP's solver. The two trit planes then fold losslessly into one 4-bit code plane that we make the persistent serving representation: disk bytes, expert-cache bytes, and kernel input are the same 4.0625-bits/weight blocks, consumed in one integer dot pass. For this conjunction (ratio-3 nine-level code, CPU-SIMD kernels, SSD expert streaming, identical persistent bytes) we likewise found no precedent. We apply this to the routed experts of DeepSeek-V4-Flash-0731, a 284B-A13B mixture-of-experts model, quantizing in one shot from the released MXFP4 expert weights and streaming experts from SSD on a 64 GB laptop. Against a 4.5-bit Q4_K baseline, measured one process per fixture with an expert-lossless anchor arm as reference control, the tied model matches the official serving API on 5/5 fixtures at step 0 (Q4_K: 4/5) and 12/14 captured continuation steps (11/14), scores 86 vs. 84 on a 100-item MMLU subset, decodes 6.7% faster in decode phase, and ships 9% smaller files: no detected fidelity difference at these small evaluation sizes, and every fixture-level difference between the arms traces to a single measured near-tie cell. The tied fit nevertheless shows higher weight-reconstruction error and worse perplexity, a measured dissociation between proxy metrics and reference fidelity. A cumulative trunk-ternarization ladder and bitwise-pinned aarch64/x86-64 kernels complete the report. All code, formats, and evaluation artifacts are open source in the fucina inference stack.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in MoE large language models, one of 12 directions in Mixture-of-experts. Builds on 21 of its 32 references that are in the map. First posted on arXiv 9 Aug 2026.

Builds on

21 of 32 references.

Closest ancestors · 4

Same direction, newest first

  1. 2026DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence26 Apr
  2. 2024QuantMoE-Bench: Examining Post-Training Quantization for Mixture-of-Experts12 Jun
  3. 2023QMoE: Practical Sub-1-Bit Compression of Trillion-Parameter Models25 Oct
  4. 2023Mixture of Quantized Experts (MoQE): Complementary Effect of Low-bit Quantization and Robustness3 Oct

Other references · 17

  1. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  2. 2016Pointer Sentinel Mixture Modelscited by 2,992
  3. 2024QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebookscited by 354
  4. 2022LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Modelscited by 98
  5. 2024HOBBIT: A Mixed Precision Expert Offloading System for Fast MoE Inferencecited by 53
  6. 2024T-MAC: CPU Renaissance via Table Lookup for Low-Bit LLM Deployment on Edgecited by 47
  7. 2025GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidancecited by 28
  8. 2024Accuracy is Not All You Needcited by 27
  9. 2025PTQTP: Post-Training Quantization to Trit-Planes for Large Language Modelscited by 9
  10. 2020DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networkscited by 8
  11. 2025Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Modelscited by 5
  12. 2026CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMscited by 5
5 more
  1. 2023Divergent Token Metrics: Measuring degradation to prune away LLM components -- and optimize quantizationcited by 4
  2. 2026Statistically-Lossless Quantization of Large Language Modelscited by 4
  3. 2026Attend to Your Own Thoughts: Breaking the Barrier for Post-Training Quantization of Reasoning LLMs through the Lens of 1.58-Bit Quantizationcited by 3
  4. 2026Displacement Is Not Direction: Evaluating Fidelity Metrics for Quantized LLM Deploymentcited by 2
  5. 2026BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Route-Block Membership Selects Packed-AWQ Arithmetic: A Controlled Single-Fixture Mechanism Study
  2. 2026MonoMoE: An Efficient Fused Mega-kernel for Quantized MoE Decoding
  3. 2026A Case Study of Selected PTQ Baselines for Reasoning LLMs on Ascend NPU
  4. 2026ButterflyMoE: Compression-Scalable Ternary Experts via Structured Butterfly Orbits
  5. 2025Dynamic Expert Quantization for Scalable Mixture-of-Experts Inference
  6. 2026CurveTQ: Rotation-Free Trellis Quantization of LLM Weights via Curvature-Weighted Search
  7. 2026LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss
  8. 2026RotaryQuant: Fitting 120B MoE Models on Consumer Hardware via Fused Compressed-Space Attention