Pith. sign in

REVIEW 4 major objections 4 minor 33 references

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

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Tying PTQTP's two ternary scales at ratio 3 yields a uniform nine-level quantizer that shows no detected fidelity difference from a 4.5-bit baseline in a disk-streamed mixture-of-experts deployment.

desk verdict Honest, open, and mechanically sound, but the fidelity-parity headline is a well-disclosed non-rejection on five prompts with two unstable cells, so treat it as a promising systems result rather than proof that tied ternary replaces q4_k. read the letter →

arxiv 2608.08910 v1 pith:UDEAH2EE submitted 2026-08-09 cs.CL cs.LG

Matteo Grella This is my paper · ORCID
classification cs.CLcs.LG
keywords ternaryquantizationnine-levelquantizerpost-trainingmixture-of-expertspersistentservingformatdiskstreamingreferencefidelityPTQTP
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PTQTP decomposes each weight group into two ternary planes with two free scales. This paper constrains those scales to a fixed ratio of three, which collapses the decomposition into a single uniform nine-level quantizer at 4.0625 bits per weight, and it makes that folded code the persistent serving representation for disk-streamed mixture-of-experts inference: disk bytes, cache bytes, and kernel input are identical, consumed in one integer dot pass. Applied to the routed experts of a 284B-parameter MoE, the tied quantizer matched the official serving API on 5/5 step-0 fixtures and 12/14 continuation steps, versus 4/5 and 11/14 for a conventional 4.5-bit quantizer, scored 86 vs 84 on a 100-item MMLU subset, decoded 6.7% faster, and shipped 9% smaller files. Because every fixture-level difference traces to one measured near-tie cell, the paper claims no detected fidelity difference, not superiority. The same measurements show higher weight-reconstruction error and worse perplexity for the tied fit, a dissociation the paper documents between proxy metrics and behavioral reference fidelity.

What carries the argument

The central mechanism is the ratio-3 scale tie $\alpha=(3s,s)$ inside PTQTP's alternating solver. With this constraint the composite code $c=3t_1+t_2$ becomes a uniformly spaced nine-level grid $\{-4,\ldots,4\}$ with one scale $s$, equivalent to fitting a uniform nine-level quantizer directly. The two ternary planes then fold losslessly into one 4-bit code (two codes per byte, four f16 column scales per 520-byte block), and because the folded code is the persistent representation, expert misses are single contiguous reads, cache tiers can use file-identical bytes, and the kernel does one integer dot pass with arithmetic decoding on NEON and AVX2/AVX-VNNI that is pinned bitwise-identical across ISAs.

What would settle it

Take a larger set of, say, 500 prompts with greedy continuations captured from the official serving API, run both quantizers in one fresh process per prompt, and count step-0 agreement and continuation prefix depth against a prespecified equivalence margin; if the tied format's agreement rate falls below the conventional 4.5-bit baseline by more than that margin, the paper's no-detected-difference claim does not hold beyond its small evaluation set.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the tie $\alpha_1 = 3s,\ \alpha_2 = s$ turns PTQTP's free two-scale ternary decomposition into a single uniform nine-level quantizer $\hat{W}=s\,c$ with $c=3t_1+t_2\in\{-4,\ldots,4\}$, and that this code can be folded into a persistent 4-bit plane (4.0625 bits/weight) that is the exact bytes on disk, in the expert-cache slab, and at the integer dot-product kernel's input. In the measured deployment, that format matched the official serving API on 5/5 step-0 fixtures and 12/14 continuation steps, versus 4/5 and 11/14 for the conventional 4.5-bit baseline, scored 86 versus 84 on a 100-item MMLU subset, decoded 6.7% faster, and shipped 9% smaller files; all fixture-level differences between the arms collapse to a single knife-edge cell, so the claim is no detected fidelity difference at these evaluation sizes, not superiority. The tied fit has higher weight-reconstruction error and worse WikiText-2 perplexity than the baseline, which the paper reads as evidence that proxy objectives and reference fidelity can disagree.

Load-bearing premise

The load-bearing premise is that the few captured greedy responses from the official serving API are a stable and meaningful reference, and that the one fixture whose first token flips under process history and tiny binary differences is measurement noise rather than a real difference between the two quantizers.

Editorial extensions

If this is right

  • The tied nine-level code can replace a conventional 4.5-bit quantizer for routed MoE experts without detected behavioral change on the tested fixtures, at a lower bit rate.
  • Persistent folded bytes make an expert miss a single contiguous read, and cached bytes are identical to disk bytes, so outputs are independent of where an expert was cached.
  • The measured dissociation between perplexity or reconstruction error and reference fidelity implies that a quantizer can look worse on proxy metrics yet match reference behavior on tested prompts, so deployment choices need behavioral anchoring.
  • The cumulative ternarization ladder shows that read-side attention projections can be ternarized without reducing fixture agreement, localizing full-model sensitivity to other components.
  • Nine percent smaller expert files at equal fidelity translate to less disk traffic and more resident experts under a fixed RAM budget.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the no-detected-difference result extends to larger, stratified reference captures, the ratio-3 tie could become a default constraint for post-training quantization of MoE experts, not a special case.
  • Editorial inference: the paper's finding that perplexity can rank arms opposite to reference agreement suggests that near-baseline quantizer comparisons should report small behavioral fixtures and task agreement alongside perplexity; this could be tested by applying the same protocol to other models and quantizer families.
  • Editorial inference: the knife-edge fixture's flip under process history and tiny binary changes implies that some argmax comparisons on tiny prompt sets are measurement noise rather than quality signals; a natural extension is to repeat near-tie fixtures many times in isolated processes and report flip probabilities.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper imposes a ratio-3 scale tie on PTQTP's two ternary planes, collapsing the decomposition into a uniform nine-level quantizer whose two trit planes fold losslessly into a single 4-bit code plane that serves as the persistent representation across disk, cache, and kernel. The construction is applied to the routed experts of DeepSeek-V4-Flash-0731, quantized one-shot from the released MXFP4 weights, and evaluated against a q4_k baseline and an expert-lossless anchor arm using official-API fixtures, a 100-item MMLU subset, WikiText-2 perplexity, and decode throughput on two machines. The headline result is that the tied quantizer shows no detected fidelity difference from q4_k at 4.06 bits/weight, with smaller files and faster decode, with every fixture-level difference traced to a single measured near-tie cell. The paper is unusually candid, explicitly labeling the result a non-rejection at small evaluation sizes and disclosing two protocol-sensitive cells whose root causes remain unexplained.

Significance. If the empirical claims hold, the contribution is practically significant: it demonstrates that a uniform nine-level quantizer can match a 4.5-bit K-quant baseline in behavioral fidelity at 4.0625 bits/weight, while the persistent folded format simplifies SSD-streamed MoE serving by keeping disk bytes, cache bytes, and kernel input identical. The paper also contributes a measured dissociation between proxy metrics (reconstruction error, perplexity) and reference fidelity, and a cumulative trunk-ternarization ladder. The manuscript is exceptionally transparent: it releases code and artifacts, pins the two ISA kernel arms bitwise-identical, uses an expert-lossless anchor arm, and discloses the exact cells that are unstable or within noise. The core tie identity is elementary arithmetic, so the contribution is in the application, measurement, and serving-system design rather than in a new mathematical result. The significance is conditional: the central empirical claim currently rests on a small, non-random fixture set with two unresolved condition-sensitive cells, and the speed claim lacks formal uncertainty and causal attribution.

major comments (4)
  1. [§4.1/Table 1, Appendix A/Table 4] The matched-depth margin favoring tied ternary (12/14 vs. 11/14) is not stable evidence. Appendix A discloses that q4_k's long memory archive depth measured 4/4 in an earlier full-process run on a pre-merge build but 2/4 deterministically under the canonical one-process-per-fixture protocol, with root causes unexplained. If the earlier value is the more faithful measurement, q4_k's depth becomes 13/14, which exceeds tied ternary's 12/14 and reverses the direction of the only continuation-depth difference. Since the abstract and §4.1 present the tied format as a practical replacement for q4_k on this deployment, this unresolved protocol swing is load-bearing. Please report both protocol values for all affected arms, add a sensitivity table that excludes both condition-sensitive cells, and either resolve the root cause or explicitly condition the replacement recommendation on the canonical protocol.
  2. [§4.1 and §6] The claim of 'no detected fidelity difference' is a non-rejection, not a demonstration of equivalence. The step-0 margin is a single knife-edge cell that flips with process history and ulp-level build changes; excluding it, the two arms are identical (4/4 at step 0, 11/13 on matched depth). With n=5 prompts, 14 dependent continuation steps, an MMLU comparison whose paired test gives p=0.6875, no prespecified equivalence margin, and a hosted reference API that can change over time, the evidence does not establish that the tied quantizer can replace q4_k without fidelity loss. The paper's wording in §6 is appropriately cautious, but the abstract and §4.1 should either state a prespecified equivalence margin with a power analysis or explicitly downgrade the headline to a descriptive non-rejection and remove any replacement implication.
  3. [§4.4/Table 3] The +6.7% decode-throughput advantage is headline material but lacks formal uncertainty and causal attribution. The paper itself notes that because the formats generate different continuations, their routed-expert workloads differ, and the speed comparison does not causally separate format, kernel, cache, and workload effects. The byte-read and pin-count decomposition is a strong mechanical check and is invariant across all rounds, but the throughput headline rests on one prompt at 32 generated tokens, six gated rounds, and includes an unresolved q4_k round at 1.95 tok/s with degraded expert-path timing. Please report confidence intervals or bootstrap uncertainty for the throughput comparison, and ideally add a workload-matched control (for example, forced decoding of identical continuation tokens) so the format-level claim is not confounded by the different continuations.
  4. [§2.1 and §4.2] The free-scale solve-log distribution is not released, and the paper's inversion finding (free scales improve perplexity while behavioral evidence is inconclusive) depends on the quality of those free-scale fits. The manuscript flags this gap, but because the inversion is a central empirical observation about the tie's cost, the absence of the solve logs and the per-expert error distribution prevents an independent check of whether the free-fit perplexity advantage is a solver artifact or a genuine property of the constraint. Please release the solve-log distribution or a representative subset, and state explicitly which of the inversion claims can be verified from the released artifacts.
minor comments (4)
  1. [Abstract] The abstract contains a typographical error: 'uniformnine-level' should read 'uniform nine-level'; also, 'q4 k' is inconsistently spaced throughout the manuscript and should be normalized to 'q4_k'.
  2. [§3, Speed protocol] The quiet-substrate acceptance gate (first attempt whose per-arm three-round spread is ≤5%) is a selection procedure that could bias the reported speed comparison; the paper states the rule was fixed in advance but not publicly registered. Please consider pre-registering the rule in the repository and reporting the number of rejected attempts and their values.
  3. [Table 2, R3 row] The caption notes that R3 was measured under the earlier protocol, but the table cell for the long memory archive (0/4) is particularly wide; it would help to mark R3's row visually with a footnote symbol in the table body rather than only in the caption.
  4. [§4.5] The cross-ISA perplexity drift (4.44 vs. 4.39 at 512 tokens) is reported as a 1.1% relative difference; restating it as an absolute NLL-space difference of 0.0113 nats/token in the main text would make the magnitude clearer and align with the paper's own guidance to state NLL-space figures when comparing small differences.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the tie is arithmetic, the quantizer is fit to released weights, and fidelity is measured against the external official API.

full rationale

The only derivation-like claim in the paper is in §2.1: constraining α=(3s,s) makes c=3t1+t2 a uniform nine-level code. That is a direct algebraic identity, with cited prior instances (patents, DBQ, balanced-ternary work), and the paper explicitly disclaims novelty of the identity itself (§5, §6). The quantizer is fitted to the released MXFP4 expert weights by an alternating solver, with no target fidelity quantity used in the fitting objective; the fidelity comparison is against the model creator's official API and a byte-exact MXFP4 anchor arm, i.e., an external reference. The 'no detected fidelity difference' claim is explicitly labeled a non-rejection on a small non-random set (§4.1, §6), so it is not presented as a derived theorem. Citations to ds4 and llama.cpp are provenance for fixtures and conversion lineage, not load-bearing support of the paper's own results, and the authors do not rely on a self-citation chain or an imported uniqueness theorem. The disclosed protocol-sensitive cells and the unresolved q4_k depth swing are measurement-validity concerns for the empirical claim, not circularity: they do not make the conclusion equal to its inputs. Hence no circular step is exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one fitted per-block scale, an arithmetic identity, and several domain assumptions about the checkpoint, the API reference, and the trunk control. No new physical or theoretical entities are posited; the tq20fx4 format is an engineering artifact, not an invented entity in the theoretical sense.

free parameters (1)
  • per-block scale s = f16 scale per 256-element block, values not enumerated in the paper
    The tied quantizer W approximately equals s times a nine-level code, with one free scale per weight group fitted to the released MXFP4 weights by the alternating solver; all fidelity and speed results depend on this fit.
assumptions (4)
  • standard math With alpha = (3s, s), the composite c = 3*t1 + t2 takes nine uniformly spaced values in {-4, ..., 4}.
    Invoked in §2.1; this is elementary arithmetic, not an empirical claim.
  • domain assumption The released DeepSeek-V4-Flash-0731 expert weights are MXFP4 and exactly dequantizable, serving as both the quantization source and the lossless reference arm.
    §2.3 and Limitations; relies on the OCP MX specification and checkpoint provenance, not independently verified here.
  • domain assumption Official-API greedy captures from ds4 are a stable reference for fidelity, and the knife-edge fixture's instability is measurement noise rather than a real arm difference.
    §3 and Appendix A; this assumption underlies the "no detected difference" conclusion.
  • domain assumption The Q8_0 trunk conversion contributes negligible error on these fixtures, inferred from the anchor arm, but no unquantized-trunk control exists.
    §2.3 and §4.1; the anchor arm shows no deficit but cannot bound trunk error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tied Trit-Planes: Constraining PTQTP to a Uniform Nine-Level Quantizer, with a Persistent Folded Format for Disk-Streamed Mixture-of-Experts Serving." pith.science (2026). https://pith.science/paper/UDEAH2EE

@misc{pith2026260808910,
  author       = {Pith},
  title        = {Pith review of: Tied Trit-Planes: Constraining PTQTP to a Uniform Nine-Level Quantizer, with a Persistent Folded Format for Disk-Streamed Mixture-of-Experts Serving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDEAH2EE}},
  note         = {Machine review of arXiv:2608.08910}
}
read the original abstract

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.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 18 canonical work pages

  1. [1]

    H. Xiao, R. Yang, Q. Yang, W. Xu, Z. Li, Y. Su, Z. Liu, H. Yang, and N. Wong. PTQTP: Post-Training Quantization to Trit-Planes for Large Language Models. arXiv:2509.16989, 2025

    arXiv 2025

  2. [2]

    E. A. Sather and S. L. Teig. Executing replicated neural network layers on inference circuit. US Patent 11,995,533 B1 (Perceive Corporation; filed Nov. 2019, priority Dec. 2018, granted May 2024)

    work page 2019

  3. [3]

    DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networks

    H. Dbouk, H. Sanghvi, M. Mehendale, and N. Shanbhag. DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networks. ECCV 2020; arXiv:2007.09818

    work page Pith review arXiv 2020

  4. [4]

    tritllm: balanced-ternary quantization for LLMs — paper (rev

    Entrit. tritllm: balanced-ternary quantization for LLMs — paper (rev. 107da27), codec (rev. 6c2b514, quantize model v2.py), uniform-d2 model cards (e.g. Qwen2.5-72B, rev. 61a31e9), and CUDA kernel (rev.7c251e6). Hugging Face; all revisions accessed 2026-08-09

    work page 2026

  5. [5]

    G. Park, B. Park, M. Kim, S. Lee, J. Kim, B. Kwon, S. J. Kwon, B. Kim, Y. Lee, and D. Lee. LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models. ICLR 2024; arXiv:2206.09557

    arXiv 2024

  6. [6]

    S. Park, J. Bae, B. Kwon, M. Kim, B. Kim, S. J. Kwon, U. Kang, and D. Lee. Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models. ACL 2025; arXiv:2506.03781

    arXiv 2025

  7. [7]

    J. Chen, J. Li, J. Xiong, W. Wang, Q. Yang, H. Xiao, Z. Li, T. Wu, M. Chen, Z. Peng, C. Tao, L. Shi, H. Yang, and N. Wong. BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models. arXiv:2602.04163, 2026

    work page Pith review arXiv 2026

  8. [8]

    Chethan Reddy G. P. ExTernD: Expanded-Rank Ternary Decomposition — Ternary LLM Post- Training Quantization with Accuracy Approaching Any Quantization Level. arXiv:2607.13511, 2026

    work page Pith review arXiv 2026

Show all 33 references
  1. [9]

    S. Wang, C. Li, Y. Kang, J. Fan, and A. Yao. Attend to Your Own Thoughts: Breaking the Barrier for Post-Training Quantization of Reasoning LLMs through the Lens of 1.58-Bit Quantization. arXiv:2608.01078, 2026

    2026 arXiv

  2. [10]

    Y. J. Kim, R. Fahim, and H. Hassan Awadalla. Mixture of Quantized Experts (MoQE): Complementary Effect of Low-bit Quantization and Robustness. NeurIPS 2023 ENLSP Workshop; arXiv:2310.02410

    2023 arXiv

  3. [11]

    Frantar and D

    E. Frantar and D. Alistarh. QMoE: Practical Sub-1-Bit Compression of Trillion-Parameter Models. MLSys 2024; arXiv:2310.16795

    2024 arXiv

  4. [12]

    P. Li, X. Jin, Z. Tan, Y. Cheng, and T. Chen. QuantMoE-Bench: Examining Post-Training Quantization for Mixture-of-Experts. arXiv:2406.08155, 2024

    2024 arXiv

  5. [13]

    J. Kim, M. El Halabi, W. Park, C. J. S. Schaefer, D. Lee, Y. Park, J. W. Lee, and H. O. Song. GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidance. ICML 2025; arXiv:2505.07004

    2025

  6. [14]

    Tseng, J

    A. Tseng, J. Chee, Q. Sun, V. Kuleshov, and C. De Sa. QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks. ICML 2024; arXiv:2402.04396. 13

    2024 arXiv

  7. [15]

    J. Wei, S. Cao, T. Cao, L. Ma, L. Wang, Y. Zhang, and M. Yang. T-MAC: CPU Renaissance via Table Lookup for Low-Bit LLM Deployment on Edge. EuroSys 2025; arXiv:2407.00088

    2025 arXiv

  8. [16]

    P. Tang, J. Liu, X. Hou, Y. Pu, J. Wang, P.-A. Heng, C. Li, and M. Guo. HOBBIT: A Mixed-Precision Expert Offloading System for Fast MoE Inference. arXiv:2411.01433, 2024

    2024 arXiv

  9. [17]

    Hashimoto

    Y. Hashimoto. Weight data compression method, weight data decompression method, weight data compression device, and weight data decompression device. US Patent 11,700,014 B2 (Socionext Inc.; priority Apr. 2021, granted Jul. 2023)

    2021

  10. [18]

    E. A. Sather, S. L. Teig, and A. F. Drimbarean. Loss-aware replication of neural network layers. US Patent 11,847,567 B1 (Perceive Corporation; priority Dec. 2018, granted Dec. 2023)

    2018

  11. [19]

    Helcig, E

    M. Helcig, E. Kurtic, and D. Alistarh. Statistically-Lossless Quantization of Large Language Models. arXiv:2605.02404, 2026

    2026 arXiv

  12. [20]

    Nikoli´ c, A

    M. Nikoli´ c, A. Hadi Zadeh, E. Torres Sanchez, and A. Moshovos. Displacement Is Not Direction: Evaluating Fidelity Metrics for Quantized LLM Deployment. arXiv:2606.19558, 2026

    2026 arXiv

  13. [21]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

    DeepSeek-AI. DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv:2606.19348, 2026

    2026

  14. [22]

    vLLM-Moet (repository; descriptive: DeepSeek-V4-class MoE serving on SM120 GPUs with persistent 2-bit expert packs over NVMe and pinned RAM)

    kacper-daftcode. vLLM-Moet (repository; descriptive: DeepSeek-V4-class MoE serving on SM120 GPUs with persistent 2-bit expert packs over NVMe and pinned RAM). GitHub repositorykacper-daftcode/vLLM-Moet, revision0a927ea; accessed 2026-08-09

    2026

  15. [23]

    Dutta, S

    A. Dutta, S. Krishnan, N. Kwatra, and R. Ramjee. Accuracy is Not All You Need. NeurIPS 2024; arXiv:2407.09141

    2024 arXiv

  16. [24]

    Deiseroth, M

    B. Deiseroth, M. Meuer, N. Gritsch, C. Eichenberg, P. Schramowski, M. Aßenmacher, and K. Kersting. Divergent Token Metrics: Measuring Degradation to Prune Away LLM Components — and Optimize Quantization. arXiv:2311.01544, 2024. NAACL 2024, pages 6764–6783, DOI 10.18653/v1/2024...

    2024 arXiv

  17. [25]

    S. Wang, C. Li, Y. Kang, J. Fan, and A. Yao. CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs. arXiv:2606.26650, 2026. ICML 2026 (Oral), PMLR 306

    2026 arXiv

  18. [26]

    Sanfilippo

    S. Sanfilippo. ds4 (DwarfStar): a DeepSeek-V4 / GLM-5.2 native inference engine, with official-API test vectors and quality harness. https://github.com/antirez/ds4 (revision b0309611, accessed Aug. 2026)

    2026

  19. [27]

    DeepSeek-V4-Flash-0731 (revision 7872f01b)

    DeepSeek-AI. DeepSeek-V4-Flash-0731 (revision 7872f01b). https://huggingface.co/ deepseek-ai/DeepSeek-V4-Flash-0731, 2026

    2026

  20. [28]

    Gerganov and contributors

    G. Gerganov and contributors. llama.cpp: LLM inference in C/C++ (GGUF and K-quant formats).https://github.com/ggml-org/llama.cpp

  21. [29]

    Merity, C

    S. Merity, C. Xiong, J. Bradbury, and R. Socher. Pointer Sentinel Mixture Models. ICLR 2017 (preprint arXiv:1609.07843, 2016)

    2017 arXiv

  22. [30]

    Janardhan

    M. Janardhan. TurboQuant-MLX: Extreme Weight and KV Cache Compression for LLMs on Apple Silicon. GitHub repository manjunathshiva/turboquant-mlx, revision a300784, Apache-2.0; accessed 2026-08-09. Includes disk-streamed MoE experts and base-3-packed ternary expert execution on...

    2026

  23. [31]

    G. Cull. DeepSeek-V4-Flash — single-workstation recipe & benchmark study. GitHub gist 30a195c(user RockmSockmJesus), revision9112837, 2026-07-22; accessed 2026-08-09

    2026

  24. [32]

    OCP Microscaling Formats (MX) Specification, Version 1.0

    Open Compute Project. OCP Microscaling Formats (MX) Specification, Version 1.0. September 2023. Defines the FP4-E2M1 element and E8M0 scale encodings. opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf

    2023

  25. [33]

    Hendrycks, C

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt. Measuring Massive Multitask Language Understanding. ICLR 2021; arXiv:2009.03300. 15

    2021 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.