Pith. sign in

REVIEW 2 major objections 5 minor 30 references

Mixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model

T0 review · 2 major / 5 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 A frozen aligner scaffold lets independently trained Transformer depth blocks be recomposed into a working language model, and a quality-parity schedule matches the monolithic baseline's reported perplexity.

desk verdict A genuinely useful proof-of-mechanism paper that shows independently trained depth slices can be recomposed into a working language model; the quality-parity headline, however, is a fitted point estimate with no equal-budget baseline and no seed variance, so the efficiency claims should be read as provisional. read the letter →

arxiv 2608.13277 v1 pith:PMIJFM6K submitted 2026-08-13 cs.CL cs.AI

Mohammed Sabry , Sean Augenstein , Keith Rush , Lucio Dery This is my paper · ORCID
classification cs.CLcs.AI
keywords mixtureoftrainingscaffoldedpretrainingmodularlanguagemodelsmodelrecompositionalignerscaffolddepthslicingefficiency
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

The paper asks whether language-model pretraining has to be a single monolithic run, or whether it can be split into smaller, independently trainable jobs that are later joined into the full model. It introduces Mixture of Training (MoT), which trains each contiguous block of a target Transformer inside a frozen pretrained aligner scaffold, then discards the scaffold and recomposes the blocks. On a 12-layer, 1.3B-parameter Gemma-style model trained on C4, the paper reports that the recomposed model reaches the same validation perplexity (15.0) as the monolithic baseline when submodel training is extended to 75k updates and followed by a 30k-step adaptation pass. The result is presented as a small-scale proof of mechanism, not a general replacement for monolithic pretraining.

What carries the argument

The load-bearing object is the frozen pretrained aligner $A = a_K \circ \cdots \circ a_1$, a shape-compatible Transformer of the same width, attention-head dimensionality, feed-forward width, token-embedding space, and output head as the target. It is sliced into $K$ blocks, and each target block $f_i$ is trained inside $S_i$, with all aligner slices frozen, so every block learns against the same representational context but no gradients are exchanged between target blocks. The aligner's role is to provide a stable interface during independent training; after Stage 1 it is discarded and the recomposed blocks are optionally adapted end-to-end.

What would settle it

Train a 12-layer target without any aligner on a different tokenizer, model width, or data stream and check whether cold-composed perplexity stays near the baseline; the paper reports 38.9 under shared data, so a configuration where no-aligner cold composition matches baseline would refute the claim that the aligner is essential. Alternatively, measure CKA or SVCCA similarity between the frozen aligner's hidden states and the trained slices' hidden states at each recomposition boundary; if similarity is no higher than between random slices, the aligner is not acting as a stable interface.

Watch

Extended reading notes

Core claim

The central claim is that a target Transformer $F = f_K \circ \cdots \circ f_1$ can be trained as $K$ independent submodels, each optimized inside a frozen aligner scaffold $S_i = a_K \circ \cdots \circ a_{i+1} \circ f_i \circ a_{i-1} \circ \cdots \circ a_1$, and that the trained blocks can then be recomposed into a recomposed model $\hat{F}$ that behaves like a coherent language model. The paper shows that a quality-parity schedule reaches perplexity 15.0, identical to the monolithic baseline, after extending Stage-1 training to 75k updates and adding a 30k-update adaptation pass, while processing 47.1B aggregate tokens and costing 285.0 EFLOPs when the aligner is fully charged to a single run. Without the aligner, cold-composition perplexity degrades from 19.3 to 38.9.

Load-bearing premise

The frozen scaffold must give each independently trained slice a shared way of representing its inputs and outputs; if that shared interface stops working outside this one 4-layer, C4 setup, the slices will not snap back together.

Editorial extensions

If this is right

  • Cold composition alone produces a usable model (PPL 19.3) at 157.9 fully charged EFLOPs, about 41% below the baseline budget, but with a clear quality gap.
  • A short 15k-step adaptation pass closes most of the gap, reaching PPL 15.9 at 189.4 fully charged EFLOPs.
  • The quality-parity schedule matches the baseline PPL of 15.0, but only when the one-time aligner cost is amortized over at least three independent target-model runs ($R \ge 3$); fully charged to one run it costs 285.0 EFLOPs versus 268.4 for the baseline.
  • Disjoint data streams for the submodels improve cold-composition quality when the aligner is present (PPL 20.3 to 19.3), but worsen it without an aligner (PPL 38.9 to 50.4).
  • Increasing the number of splits from 2 to 4 cuts Stage-1 compute by roughly 37% but degrades cold-composition perplexity (PPL 19.3 to 24.8 under the 4-layer aligner).

Reading between the lines

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

  • If the aligner interface transfers to other architectures, depths, and data streams, pretraining could become a library of reusable depth blocks whose interfaces are fixed by a shared scaffold, enabling mix-and-match assembly of new models without joint training.
  • The fault-tolerance benefit the paper sketches is untested: because each scaffold job is independently restartable, a hardware failure would in principle only redo one slice, but recovery cost depends on how much of the aligner and data pipeline must be re-run.
  • A direct test of interface stability—measuring hidden-state similarity (e.g., CKA) across recomposition boundaries before and after adaptation—would likely predict which block pairs compose well without needing a full training run.
  • The same scaffolded scheme could be combined with data specialization: training different slices on different corpora or objectives (e.g., code vs. natural language) might yield recomposed models with controllable skill mixtures, if the aligner keeps the interface stable.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces Mixture of Training (MoT), a scaffolded modular pretraining procedure. A target Transformer is partitioned into K contiguous layer blocks; each block f_i is trained independently (no gradient exchange) inside a frozen pretrained 'aligner' scaffold S_i that sandwiches the block with shape-compatible aligner slices, after which the blocks are recomposed into F and optionally given a short end-to-end adaptation pass. Experiments on a 12-layer, 1.3B-parameter Gemma-style model on C4 report: cold composition reaches PPL 19.3 versus 15.0 for the 128k-step monolithic baseline (268.4 EFLOPs); a 15k-step adaptation pass reaches PPL 15.9 at 189.4 fully-charged EFLOPs; and a 'quality-parity' schedule (75k scaffolded updates plus 30k adaptation, 47.1B aggregate tokens, 255.3 train EFLOPs / 285.0 fully charged) reaches PPL 15.0. Ablations show the aligner is essential (no-aligner cold PPL 38.9), disjoint submodel data streams help under the tested aligner (19.3 vs 20.3), and K=4 trades quality for compute. The paper provides explicit FLOP, token, memory, and idealized critical-path accounting, gives an amortization formula (effective cost 255.3 + 29.7/R per run), and explicitly positions the work as a small-scale proof of mechanism rather than a general replacement for monolithic pretraining.

Significance. The paper's strengths are real. I checked the arithmetic in Appendices B and D; the FLOP, token, and critical-path numbers are internally consistent (e.g., 47.1B tokens = 75k×2×256×1024 + 30k×256×1024; 128k/(75×0.61 + 30) ≈ 1.7×). The cold-composition result together with the no-aligner control (19.3 vs 38.9) is a clean behavioral demonstration that independently trained depth slices can be recomposed when trained inside a shared scaffold. The accounting is unusually transparent: Section 4 explicitly concedes the absence of compute-matched controls, measured wall-clock times, downstream evaluations, and direct interface diagnostics, and Appendix E discloses that main results are individual runs. If the composability result survives seed variance and transfers beyond this single 1.3B/C4 instance, MoT opens a genuinely useful design space of parallel, restartable, and potentially reusable scaffolded sub-runs.

major comments (2)
  1. [§3 Table 1; Appendix E] The quality-parity claim (PPL 15.0 in Table 1) rests on a single unseeded run, as Appendix E states ('The main results are individual runs'), and the paper does not document how the 75k + 30k schedule was chosen. If the schedule was selected so that the final perplexity lands on the baseline value of 15.0, the parity result is a fitted operating point rather than an independent prediction, and with typical run-to-run variance the 15.0-vs-15.0 equality could be within noise. This matters because both the abstract's headline claim ('a quality-parity schedule reaches the same reported perplexity as the monolithic baseline') and the amortization break-even (R≥3) rest on this single point. Please report at least three seeds for the quality-parity and cold-composition schedules with mean and spread, or explicitly re-label the result as a single achieved operating point and document the schedule-selection procedure.
  2. [§3 Table 1; §4] No equal-resource monolithic control is reported. The fully charged MoT parity schedule (285.0 EFLOPs, 47.1B tokens) is compared only against the 128k-step baseline (268.4 EFLOPs, 33.6B tokens); a monolithic run at the MoT FLOP budget (~36B tokens, ≈136k steps) or at the MoT token budget (~375 EFLOPs) would determine whether MoT matches, beats, or loses to monolithic training at equal resources, and neither is reported. Section 4 concedes this. The omission is load-bearing because the R≥3 amortization argument is a direct function of the parity schedule's 255.3/285.0 EFLOP costs, which are not yet established as necessary-or-sufficient for parity. In addition, the 30k Stage-2 pass (23% of baseline steps, all layers trainable) closes most of the cold-composition gap (19.3 to 15.0); without a control such as a 30k-step end-to-end run from random initialization, the fraction of the parity result attributable to the scaffolded Stage-1 blocks is not quantified. I recommend adding at least one monolithic checkpoint at the ~285-EF budget and a random-init-plus-30k-adaptation control, or removing the parity framing from the abstract.
minor comments (5)
  1. [Appendix B; Table 5] Under the paper's stated 6ND accounting rule, the 4-layer aligner's 29.7 EFLOPs at a 10B-token budget implies an effective non-embedding parameter count of about 0.50B, whereas Table 5 lists 0.4B parameters; the 6ND value computed from the listed count is 24 EFLOPs, about 19% below the reported figure. A sentence reconciling this (e.g., naming any extra terms included in the aligner FLOPs) would keep the accounting fully transparent.
  2. [Appendix E; §3] The disclosure that 'the main results are individual runs' appears only in Appendix E; since Table 1 presents the headline comparisons without variance information, this caveat should also appear in the main text near Table 1.
  3. [Table 6; §F.1] The disjoint-vs-shared data effect (20.3 vs 19.3 at K=2) and the K=2-vs-K=4 comparisons are single unreplicated runs with small effect sizes; a sentence in the main text stating that these are unreplicated observations would prevent readers from over-weighting differences of about one perplexity point.
  4. [§2] In the definition S_i = a_K ∘ ... ∘ a_{i+1} ∘ f_i ∘ a_{i-1} ∘ ... ∘ a_1, the degenerate cases i=1 and i=K are not spelled out; a one-line clarification that empty aligner prefixes or suffixes are dropped would remove ambiguity.
  5. [§1; §4] The abstract and Section 1 describe the aligner as supplying a 'stable representational interface' that makes blocks compatible, while Section 4 correctly states that only behavioral evidence exists and that 'a fuller mechanistic account would require direct interface diagnostics'; the earlier phrasing should be hedged to match that stated evidentiary level.

Circularity Check

1 steps flagged · score 4.0 of 10

Quality-parity PPL match is a fitted schedule target; cold-composition composability claim is independent.

  1. fitted input called prediction [Section 3 (Table 1) and Appendix E]
    "The monolithic baseline is trained end-to-end for 128k updates, processing 33.6B tokens and reaching perplexity 15.0 at 268.4 EFLOPs. ... MoT quality parity reinvests part of the saved compute by extending submodel training to 75k updates and using a 30k adaptation pass. ... The main results are individual runs; only the scheduler ablation was replicated."

    The 'quality parity' schedule is selected to land on the baseline PPL: the paper chooses 75k Stage-1 updates plus a 30k Stage-2 pass and reports PPL 15.0, exactly the monolithic baseline's reported PPL. The equality is therefore the fitted target of schedule selection, not an independent prediction of MoT's recomposition procedure. Table 1's parity row is equivalent to the input baseline PPL by construction, and the single unseeded run means the match is a tuned point estimate rather than a derived consequence. The cold-composition row (PPL 19.3, measured before adaptation) remains an independent composability result, but the headline 'quality parity reaches the same reported perplexity' is a fitted outcome.

full rationale

The paper's derivation is mostly self-contained: FLOP, memory, and critical-path tables are arithmetic from stated assumptions in Appendices B-D, and the aligner-essential and data-stream ablations are independent empirical comparisons. There is no load-bearing self-citation: Deep Incubation and other related work are cited as prior art, not as authority for the present result. The one circular element is the quality-parity row: the schedule is named after and tuned to the baseline PPL of 15.0, so reporting 15.0 is the fitted target rather than an independent prediction. The absence of a compute-matched monolithic control and the single unseeded run further weaken the parity claim, though these are limitations rather than circularity. However, the cold-composition result (PPL 19.3, no adaptation) independently demonstrates that independently trained slices can be recomposed into a usable language model, and that conclusion does not depend on the fitted parity row. Hence the score is 4: one fitted 'prediction' contributes partial circularity, while the core composability claim retains independent support.

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

The central composability result rests on the untested premise that a frozen aligner provides a transferable representational interface, plus the standard 6ND FLOP model and an idealized critical-path model. The compute comparisons also depend on excluding output-projection costs, which the paper acknowledges could make its FLOP numbers optimistic at a 256k-token vocabulary.

free parameters (4)
  • Stage-1 submodel updates S1 = 75k for quality parity; 50k for the lower-compute schedule
    Chosen by hand to hit the target perplexity; the quality-parity schedule is tuned so the recomposed model reaches the baseline PPL of 15.0.
  • Stage-2 adaptation steps S2 = 30k for quality parity; 15k otherwise
    Selected to close the cold-composition gap; the parity schedule is fitted to reach PPL 15.0 after adaptation.
  • Aligner depth and token budget = 4 layers, 10B tokens (M=25) in the headline; M=50 and M=100 variants ablated
    A hyperparameter of the method; the headline uses the smallest aligner after ablations showed bigger aligners do not improve adapted PPL.
  • Split count K = 2
    The headline uses K=2; K=4 is ablated and shows worse cold-composition PPL.
assumptions (4)
  • domain assumption C ≈ 6ND training-FLOP approximation, layer-dominated, excluding embedding lookup and output projection
    Used in Appendix B for all compute numbers; the paper notes the 256k-vocab output head may make its FLOP values optimistic.
  • ad hoc to paper The frozen pretrained aligner provides a stable representational interface making depth slices composable
    Central premise of the method, asserted in Section 2 and Appendix A; the paper says it lacks direct interface diagnostics.
  • domain assumption Idealized critical path: equal cost per layer, concurrent Stage 1 jobs, backward cost for frozen downstream layers
    Used in Appendix D to produce the 1.7x, 2.8x, and 4.2x ratios; these are not measured equal-hardware speedups.
  • domain assumption Separately sampled C4 dataloader streams approximate disjoint data
    Appendix E states no sample-level dedup was enforced, so 'disjoint streams' are not truly disjoint corpora.
invented entities (1)
  • Aligner scaffold A = a_K o ... o a_1
    purpose: Frozen pretrained representational interface in which each target block is trained and through which blocks are made compatible at recomposition
    A methodological construct adapted from Deep Incubation's meta model; it has no falsifiable handle outside this paper, and its efficacy is the paper's own empirical result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model." pith.science (2026). https://pith.science/paper/PMIJFM6K

@misc{pith2026260813277,
  author       = {Pith},
  title        = {Pith review of: Mixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMIJFM6K}},
  note         = {Machine review of arXiv:2608.13277}
}
read the original abstract

We ask whether language-model pre-training can be decomposed into smaller, independently trainable jobs that can later be recomposed into a coherent larger model. We introduce Mixture of Training (MoT), a scaffolded modular pre-training procedure that partitions a target Transformer into contiguous layer blocks, trains each block inside a frozen pretrained aligner scaffold, and then recomposes the trained blocks with an optional short end-to-end adaptation pass. On a 1.3B-parameter Gemma-style model trained on C4, MoT provides a small-scale proof of mechanism: independently trained depth slices can be recomposed into a usable language model, and a quality-parity schedule reaches the same reported perplexity as the monolithic baseline. This parity setting processes more aggregate tokens and has a shorter idealized layer-equivalent critical path after aligner preparation; its effective compute advantage depends on reusing the aligner across runs. We therefore present MoT not as a general replacement for monolithic pre-training, but as a small-scale framework for studying whether scaffolded sub-runs can act as reusable training units.

Figures

Figures reproduced from arXiv: 2608.13277 by the authors.

Figure 1
Figure 1. Mixture of Training (MoT). The illustrated preparation and slicing steps form Stage 0; scaffolded block training is Stage 1; and recomposition with optional end-to-end adaptation is Stage 2. Only the target block in each scaffold is updated. We report perplexity, the exponential of mean token-level cross-entropy on held-out C4. Baseline and MoT runs share the data source, optimizer family, and batch setting, with st… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 10 canonical work pages

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    Brown and Benjamin Chess and Rewon Child and Scott Gray and Alec Radford and Jeffrey Wu and Dario Amodei , title =

    Jared Kaplan and Sam McCandlish and Tom Henighan and Tom B. Brown and Benjamin Chess and Rewon Child and Scott Gray and Alec Radford and Jeffrey Wu and Dario Amodei , title =. CoRR , volume =. 2020 , url =. 2001.08361 , timestamp =

    arXiv 2020

  5. [5]

    2022 , eprint=

    Training Compute-Optimal Large Language Models , author=. 2022 , eprint=

    2022

  6. [6]

    2025 , eprint=

    L4: Diagnosing Large-scale LLM Training Failures via Automated Log Analysis , author=. 2025 , eprint=

    work page 2025

  7. [7]

    2023 , eprint=

    Accelerating Large Language Model Decoding with Speculative Sampling , author=. 2023 , eprint=

    2023

  8. [8]

    2020 , eprint=

    GLU Variants Improve Transformer , author=. 2020 , eprint=

    2020

Show all 30 references
  1. [9]

    2022 , editor =

    Du, Nan and Huang, Yanping and Dai, Andrew M and Tong, Simon and Lepikhin, Dmitry and Xu, Yuanzhong and Krikun, Maxim and Zhou, Yanqi and Yu, Adams Wei and Firat, Orhan and Zoph, Barret and Fedus, Liam and Bosma, Maarten P and Zhou, Zongwei and Wang, Tao and Wang, Emma and Web...

    2022

  2. [10]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Andreas, Jacob and Rohrbach, Marcus and Darrell, Trevor and Klein, Dan , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  3. [11]

    CoRR , volume =

    Clemens Rosenbaum and Ignacio Cases and Matthew Riemer and Tim Klinger , title =. CoRR , volume =. 2019 , url =. 1904.12774 , timestamp =

    2019 arXiv

  4. [12]

    2024 , eprint=

    Gemma 2: Improving Open Language Models at a Practical Size , author=. 2024 , eprint=

    2024

  5. [13]

    2403.08295 , archivePrefix =

    Gemma: Open Models Based on Gemini Research and Technology , year =. 2403.08295 , archivePrefix =

    arXiv

  6. [14]

    Distill , year =

    Chris Olah and Nick Cammarata and Ludwig Schubert and Gabriel Goh and Michael Petrov and Shan Carter , title =. Distill , year =. doi:10.23915/distill.00024.001 , url =

    doi

  7. [15]

    2025 , eprint=

    Distillation Scaling Laws , author=. 2025 , eprint=

    2025

  8. [16]

    2022 , eprint=

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time , author=. 2022 , eprint=

    2022

  9. [17]

    bert2 BERT : Towards Reusable Pretrained Language Models

    Chen, Cheng and Yin, Yichun and Shang, Lifeng and Jiang, Xin and Qin, Yujia and Wang, Fengyu and Wang, Zhi and Chen, Xiao and Liu, Zhiyuan and Liu, Qun. bert2 BERT : Towards Reusable Pretrained Language Models. Proceedings of the 60th Annual Meeting of the Association for Comp...

    2022 doi

  10. [18]

    CoRR , volume =

    Yamini Bansal and Preetum Nakkiran and Boaz Barak , title =. CoRR , volume =. 2021 , url =. 2106.07682 , timestamp =

    2021 arXiv

  11. [19]

    2023 , eprint=

    Model Stitching: Looking For Functional Similarity Between Representations , author=. 2023 , eprint=

    2023

  12. [20]

    Efficient Large-Scale Distributed Training of Conditional Maximum Entropy Models , url =

    Mcdonald, Ryan and Mohri, Mehryar and Silberman, Nathan and Walker, Dan and Mann, Gideon , booktitle =. Efficient Large-Scale Distributed Training of Conditional Maximum Entropy Models , url =

  13. [21]

    Zhao and Kelvin Guu and Adams Wei Yu and Brian Lester and Nan Du and Andrew M

    Jason Wei and Maarten Bosma and Vincent Y. Zhao and Kelvin Guu and Adams Wei Yu and Brian Lester and Nan Du and Andrew M. Dai and Quoc V. Le , title =. CoRR , volume =. 2021 , url =. 2109.01652 , timestamp =

    2021 arXiv

  14. [22]

    2022 , eprint=

    Scaling Instruction-Finetuned Language Models , author=. 2022 , eprint=

    2022

  15. [23]

    2022 , eprint=

    Training language models to follow instructions with human feedback , author=. 2022 , eprint=

    2022

  16. [24]

    2022 , eprint=

    Constitutional AI: Harmlessness from AI Feedback , author=. 2022 , eprint=

    2022

  17. [25]

    CoRR , volume =

    Deepak Narayanan and Mohammad Shoeybi and Jared Casper and Patrick LeGresley and Mostofa Patwary and Vijay Korthikanti and Dmitri Vainbrand and Prethvi Kashinkunti and Julie Bernauer and Bryan Catanzaro and Amar Phanishayee and Matei Zaharia , title =. CoRR , volume =. 2021 , ...

    2021 arXiv

  18. [26]

    2026 , eprint=

    DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation , author=. 2026 , eprint=

    2026

  19. [27]

    The Twelfth International Conference on Learning Representations , year =

    Masked Structural Growth for 2x Faster Language Model Pre-training , author =. The Twelfth International Conference on Learning Representations , year =

  20. [28]

    Stacking Your Transformers: A Closer Look at Model Growth for Efficient

    Wenyu Du and Tongxu Luo and Zihan Qiu and Zeyu Huang and Yikang Shen and Reynold Cheng and Yike Guo and Jie Fu , booktitle =. Stacking Your Transformers: A Closer Look at Model Growth for Efficient. 2024 , doi =

    2024

  21. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Deep Incubation: Training Large Models by Divide-and-Conquering , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2023 , url =

    2023

  22. [30]

    Ka Man Lo and Yiming Liang and Wenyu Du and Yuantao Fan and Zili Wang and Wenhao Huang and Lei Ma and Jie Fu , journal =. M2M. 2024 , url =

    2024

Pith tools

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