Memory Efficient Tabular Foundation Models

# Memory Efficient Tabular Foundation Models
Source: [https://arxiv.org/html/2607.27546](https://arxiv.org/html/2607.27546)
###### Abstract

Tabular Foundation Models, such as TabPFN, have received a large amount of recent attention due to their performance on in\-context tabular machine learning tasks, which often exceeds classical baselines\. However, practical deployment considerations of these models have received less attention\. In this paper we investigate the memory requirements for these models\. We demonstrate that employing model compression approaches can enable memory reductions of up to 7\.6×\\timeswith similar levels of performance, reducing deployment requirements by nearly87%87\\%\. Our work provides insight to practitioners seeking efficient deployment of these models in practical settings\.

Machine Learning, ICML

## 1Introduction

Tabular machine learning is used heavily in practical domains such as healthcare, finance, operations, and enterprise analytics, where models are often deployed under strict constraints on latency, cost, hardware availability, and data movement\(Jianget al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib26); Shwartz\-Ziv and Armon,[2021](https://arxiv.org/html/2607.27546#bib.bib39); Borisovet al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib42)\)\. Recent tabular foundation models \(TFMs\), including the TabPFN and TabICL series, are attractive in these settings because they can make strong predictions in\-context, reducing the need for task\-specific training, hyperparameter tuning, and bespoke model\-selection pipelines\(Jianget al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib26); Hollmannet al\.,[2023](https://arxiv.org/html/2607.27546#bib.bib7); Quet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib3); Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1)\)\. This makes them appealing for organizations that need to solve many small or medium\-sized tabular prediction problems repeatedly\.

However, practical adoption depends not only on predictive performance, but also on whether these models can be deployed and communicated efficiently\(Fuet al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib44)\)\. Enterprises must contend with memory movements, hardware constraints, and practical bandwidth considerations\. These issues are well\-known constraints in servicing large language models, where memory footprint directly affects deployment cost, latency, and the feasibility of local inference\(Menghani,[2023](https://arxiv.org/html/2607.27546#bib.bib45); Jeyaraman,[2025](https://arxiv.org/html/2607.27546#bib.bib46)\)\. To address this, memory\-efficient techniques such as model quantization are now ubiquitous in practical pipelines\(Dettmerset al\.,[2023](https://arxiv.org/html/2607.27546#bib.bib18); Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\.

These memory\-based limitations also exist for TFMs, however to date most work has focused on reducingcontextconstraints, due to architectural limitations and attention\-based inference that grows quickly with sample and feature dimensions\(Zabërgjaet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib41); Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1); Quet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib4); Zenget al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib40); Liu and Ye,[2025](https://arxiv.org/html/2607.27546#bib.bib47)\), leaving practical deployment considerations relatively unexplored\.

![Refer to caption](https://arxiv.org/html/2607.27546v1/hero-figure.png)Figure 1:INT4 quantization makes Tabular Foundation Models substantially more memory\-efficient while preserving their accuracy advantage\. Quantized models have a7\.6×7\.6\\timeslower memory footprint relative to full precision baselines, with negligible decline in accuracy, remaining above the strongest tuned classical baseline\. This demonstrates the potential for quantized TFMs to achieve high predictive performance at a fraction of the deployment cost\.This paper asks the question:Can we reduce the deployment requirements for TFMs while retaining their predictive performance?We evaluate whether existing quantization methods can be employed to reduce the memory footprint of the TabPFN family of models, using standard tabular benchmarks and classical baselines for context\.

![Refer to caption](https://arxiv.org/html/2607.27546v1/ELO.png)\(a\)ELO ratings from pairwise accuracy comparisons\. Quantized TabPFN models remain close to their FP32 counterparts and rank above all tuned and untuned classical baselines\.
![Refer to caption](https://arxiv.org/html/2607.27546v1/winloss.png)\(b\)Pairwise win probabilities on evaluated datasets\. Quantized TabPFN models consistently beat classical baselines, with most losses concentrated against FP32 TabPFN variants\.

Figure 2:Quantization preserves most of the accuracy gains of Tabular Foundation Models\. Quantized TabPFN variants incur only a modest drop relative to FP32 models while maintaining a clear advantage over strong classical ML baselines\.Our results indicate that quantization is a simple and practical route to memory\-efficient TFMs\. Across 30 datasets, we find that quantizing to INT4 precision reduces memory footprint of TFMs while maintaining similar predictive performance to the full precision counterparts\. These results show that deployment needs can be reduced by up to 7\.6×\\timeswhile maintaining performance comparable to tree\-based models like CatBoost\(Prokhorenkovaet al\.,[2018](https://arxiv.org/html/2607.27546#bib.bib25)\)\.

In this paper we make the following contributions:

- •We highlight the practical need for memory\-efficient tabular machine learning models and demonstrate that large improvements can be obtained using standard quantization methods\.
- •We provide a systematic evaluation of these quantization approaches, and qualitative considerations for practical deployment\.

We demonstrate our experiments on a range of standard tabular datasets using the TabPFNv2\.5 and TabPFNv2\.6 variants from the TabPFN family of models\(Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1)\)\.

Table 1:Comparison of model performance and compression methods \(30 Datasets\)\.
## 2Related Literature

### 2\.1Tabular Foundation Models

Tabular Foundation Models are a class of pre\-trained models that can be applied to tabular machine learning problems without additional training\(Jianget al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib26)\)\. These models are trained to predict a posterior prediction \(yt​e​s​t\)y\_\{test\}\)based on a tuple of in\-context data \(xt​r​a​i​n,yt​r​a​i​n;xt​e​s​tx\_\{train\},y\_\{train\};x\_\{test\}\)\(Hollmannet al\.,[2023](https://arxiv.org/html/2607.27546#bib.bib7); Quet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib3)\)\. The training of these models involves extensive pre\-training on simulated data of causal relationships and are known as Prior Fitted Networks \(PFNs\)\(Mülleret al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib27); Hollmannet al\.,[2023](https://arxiv.org/html/2607.27546#bib.bib7)\)\. Follow up work has sought to extend these models to more realistic tabular scenarios \(e\.g\. missing data, categorical data, time\-series\), and to handle limitations in the feature and sample dimension\(Hollmannet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib6); Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1); Quet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib3),[2026](https://arxiv.org/html/2607.27546#bib.bib4); Hooet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib28)\)\. The strong performance of these models relative to classical baselines has generated applied interest in a wide range of practical domains such as medical machine learning, financial forecasting, and a variety of industrial uses\(Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1)\)\. The practical handling of large contexts has received a large amount of interest\(Zenget al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib40); Zabërgjaet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib41)\)\. However, memory requirements of weight deployment have received relatively little attention\.

### 2\.2Quantization and Model Compression

Quantization is a mapping involving the discretization of a signal\(Gersho and Gray,[1991](https://arxiv.org/html/2607.27546#bib.bib9)\)\. Within machine learning, the term is typically used to refer toweightoractivationquantization, where the goal is to reduce memory by using lower precision values\(Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\.111Note: the termquantizationis occasionally used to refer to discretization of features in tabular machine learning contexts\. We use the term to refer to weight quantization in this paper reflecting our interest in memory efficiency in Tabular Foundation Models\.We outline briefly broad trends in the literature and key design decisions\.

Early work includes binary neural networks\(Rastegariet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib11); Hubaraet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib16)\), the use of quantization aware\-training, and integer quantization\(Jacobet al\.,[2018](https://arxiv.org/html/2607.27546#bib.bib22)\)\. Classical results show that MSE optimal quantizers correspond closely to a clustering problem\(Gersho and Gray,[1991](https://arxiv.org/html/2607.27546#bib.bib9)\), motivating codebook k\-means quantizers\(Hanet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib10)\)\.

Recent advances in quantization have included the use of random rotations\(Zandiehet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib2); Liuet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib13); Ashkbooset al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib15)\); the use of singular value decomposition to reduce the effect of outlier values\(Li\*et al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib14)\); activation\-aware quantization \(AWQ\)\(Linet al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib12)\); and dedicated libraries such as GPTQ and BitsandBytes\(Dettmerset al\.,[2022a](https://arxiv.org/html/2607.27546#bib.bib20),[b](https://arxiv.org/html/2607.27546#bib.bib19),[2023](https://arxiv.org/html/2607.27546#bib.bib18); Frantaret al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib21)\)\. A large body of literature exists on applying quantization methods to reduce the memory requirements for modality specific use\-cases such as large language model compression\(Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\. However, little work has been devoted to weight quantization within TFMs\.

### 2\.3Memory Efficient Tabular Models

There is some work in reducing memory requirements for tabular machine learning models\. For example, through the use of low\-precision decision gradient boosted trees\(Herrmannet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib31); Shiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib29)\); and in applying integer quantization to enable efficient performance on FPGA\-enabled edge devices\(Alsharariet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib30)\)\. Separately, within TFMs there are some works on achieving high tabular performance with small MLP or distilled models such as TabM\(Gorishniyet al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib32)\)and TabDistill\(Dissanayake and Dutta,[2025](https://arxiv.org/html/2607.27546#bib.bib34)\)\. In addition, inference memory requirements have been reduced by compressing context\(Zabërgjaet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib41)\), or through hypernetworks\(Muelleret al\.,[2025](https://arxiv.org/html/2607.27546#bib.bib33)\)\.

## 3Method

### 3\.1Quantization Methods

We evaluate three INT4 post\-training quantization approaches indicative of recent trends in model compression: K\-Means Quantization\(Hanet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib10)\), AWQ\(Linet al\.,[2024](https://arxiv.org/html/2607.27546#bib.bib12)\), and GPTQ\(Frantaret al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib21)\)\. We focus on post\-training quantization \(PTQ\) rather than quantization\-aware training, as this can be conducted on a pre\-trained model and does not require access to the original pipeline\. For GPTQ and AWQ, originally developed for autoregressive language models, we adapt their calibration procedures to TabPFN’s in\-context learning setting by treating the\(xt​r​a​i​n,yt​r​a​i​n;xt​e​s​t\)\(x\_\{train\},y\_\{train\};x\_\{test\}\)tuple as the calibration input\. For K\-Means clustering quantization, we implement an INT4 bit\-packing scheme to reduce model artifact size on disk\.

![Refer to caption](https://arxiv.org/html/2607.27546v1/rate-distortion.png)Figure 3:Effect of K\-means quantization bit width on TabPFN accuracy and model size\. Accuracy improves sharply from 1\-bit to 2\-bit quantization and largely saturates by 4\-bit quantization, where both TabPFN\-v2\.5 and TabPFN\-v2\.6 match or approach full precision performance while requiring only 5\.4 MB\. Further increasing to FP32 yields negligible accuracy gains but increases model size substantially to 41\.0 MB, indicating that 4\-bit quantization provides the best accuracy\-memory trade\-off\.
### 3\.2Experimental Setup

We evaluate on a subset of 30 datasets from the OpenML library\(Vanschorenet al\.,[2013](https://arxiv.org/html/2607.27546#bib.bib35)\), selected to match feature and sample limitations of TabPFN\. A full list of datasets is provided in Appendix[A\.1](https://arxiv.org/html/2607.27546#A1.SS1)\. We test our three quantization methods on two TabPFN variants: TabPFNv2\.5, and TabPFNv2\.6\(Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1)\)\. Memory is evaluated following BZIP2 compression, replicating a practical deployment pipeline\(Seward,[1996](https://arxiv.org/html/2607.27546#bib.bib48)\)\. We benchmark against standard tabular machine learning methods: XGBoost\(Chen and Guestrin,[2016](https://arxiv.org/html/2607.27546#bib.bib23)\), CatBoost\(Prokhorenkovaet al\.,[2018](https://arxiv.org/html/2607.27546#bib.bib25)\), LightGBM\(Keet al\.,[2017](https://arxiv.org/html/2607.27546#bib.bib24)\), KNN\(Fix and Hodges Jr,[1951](https://arxiv.org/html/2607.27546#bib.bib37); Cover and Hart,[1967](https://arxiv.org/html/2607.27546#bib.bib38)\), and Random Forest\(Breiman,[2001](https://arxiv.org/html/2607.27546#bib.bib36)\), evaluated in both default\-configuration and tuned variants\. Hyperparameter details are provided in Appendix[A\.2](https://arxiv.org/html/2607.27546#A1.SS2.SSS0.Px1)\.

## 4Results

Table[1](https://arxiv.org/html/2607.27546#S1.T1)and Figure[1](https://arxiv.org/html/2607.27546#S1.F1)present quantization performance across TabPFNv2\.5 and TabPFNv2\.6\. All three INT4 quantization methods achieve substantial compression \(6\.0\-7\.65×\\times\) with minimal accuracy degradation \(≤0\.005\\leq 0\.005\)\. Among quantization methods, K\-means INT4 achieves the highest compression ratio \(7\.65×\\timesfor TabPFNv2\.5, and 7\.56×\\timesfor TabPFNv2\.6\) while incurring the smallest accuracy loss \(0\.001\), outperforming both GPTQ and AWQ across both model versions\. Critically, all quantized variants maintain accuracy above the best classical baseline \(LightGBM tuned: 0\.864\), demonstrating that compression preserves TabPFN’s performance advantage over traditional methods\.

Several further observations support these results\. First, a bit\-width ablation \(Table[2](https://arxiv.org/html/2607.27546#A3.T2)\) confirms INT4 as the practical optimum: 2\-bit quantization recovers only∼\\sim90% of baseline accuracy and 1\-bit collapses to near\-random performance, while INT4 retains over 99%\. Second, Figure[2](https://arxiv.org/html/2607.27546#S1.F2)shows that the quantized models retain a consistent dataset\-level advantage over classical baselines, with a marginal relative ELO loss relative to full precision versions\. Third, performance variance across the 30 datasets remains essentially unchanged between baseline and quantized variants \(see Appendix Table[3](https://arxiv.org/html/2607.27546#A3.T3)\), indicating that quantization does not introduce dataset\-dependent instability\. Finally, while deployable memory is substantially reduced, inference\-time memory and runtime are largely unchanged \(see Appendix Table[4](https://arxiv.org/html/2607.27546#A3.T4)\)\. This is consistent with TabPFN’s inference cost being dominated by attention activations rather than weights, motivating activation\-level compression as future work\.

## 5Conclusion and Limitations

In this work, we investigated the memory efficiency of TFMs, with a particular focus on quantization\-based model compression\. Our results demonstrate that substantial reductions in deployment memory are achievable with quantization without compromising predictive performance, making these models more practical for use within industry\.

#### Limitations

Our study focuses exclusively on post\-training quantization, which may be suboptimal relative to quantization\-aware training, which may produce higher performance at lower\-precision with access to the training pipeline\(Rastegariet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib11); Hanet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib10); Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\. In addition, our evaluation considers only a subset of available quantization techniques, leaving more advanced approaches such as TurboQuant\(Zandiehet al\.,[2026](https://arxiv.org/html/2607.27546#bib.bib2)\)and other emerging methods unexplored\. As is standard in quantization for memory deployment, our method requires dequantization prior to inference and so does not enable run\-time memory improvements\(Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\. We also do not incorporate kernel\-level or hardware\-aware optimizations, which may enable FPGA/edge\-device implementations\. Finally, our method retains the architectural limitations of the foundation models tested which constrain the maximum context window\(Grinsztajn and others,[2025](https://arxiv.org/html/2607.27546#bib.bib1)\)\.

#### Future Work

Several promising directions exist for extending this work\. Beyond quantization, complementary forms of model compression including pruning\(Hanet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib10)\), low\-rank adaptation \(LoRA\)\(Huet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib49)\), and knowledge distillation\(Hintonet al\.,[2015](https://arxiv.org/html/2607.27546#bib.bib50)\), may yield further gains in memory and computational efficiency\. Improving inference\-time memory efficiency represents another valuable avenue, for example by combining memory\-efficient attention mechanisms or context\-handling strategies with a quantized approach\. Exploring quantization\-aware training may further improve efficiency while preserving low\-precision predictive performance\(Rastegariet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib11); Hanet al\.,[2016](https://arxiv.org/html/2607.27546#bib.bib10)\)\. Activation quantization strategies, which enable inference in low\-precision are a further avenue\(Gholamiet al\.,[2022](https://arxiv.org/html/2607.27546#bib.bib17)\)\. Finally, integrating hardware\-aware optimizations and extending TFM models to physical edge devices remains an exciting potential future direction for a range of practical memory\-bound applications\.

## References

- M\. Alsharari, S\. T\. Mai, R\. Woods, and C\. Reaño \(2025\)Efficient integer\-only\-inference of gradient boosting decision trees on low\-power devices\.IEEE Transactions on Circuits and Systems I: Regular Papers72\(1\),pp\. 241–253\.External Links:[Document](https://dx.doi.org/10.1109/TCSI.2024.3446582)Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- S\. Ashkboos, A\. Mohtashami, M\. L\. Croci, B\. Li, P\. Cameron, M\. Jaggi, D\. Alistarh, T\. Hoefler, and J\. Hensman \(2024\)QuaRot: outlier\-free 4\-bit inference in rotated LLMs\.InThe Thirty\-eighth Annual Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=dfqsW38v1X)Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- V\. Borisov, T\. Leemann, K\. Seßler, J\. Haug, M\. Pawelczyk, and G\. Kasneci \(2024\)Deep neural networks and tabular data: a survey\.IEEE Transactions on Neural Networks and Learning Systems35\(6\),pp\. 7499–7519\.External Links:[Document](https://dx.doi.org/10.1109/TNNLS.2022.3229161)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1)\.
- L\. Breiman \(2001\)Random forests\.Machine Learning45\(1\),pp\. 5–32\.External Links:[Document](https://dx.doi.org/10.1023/A%3A1010933404324)Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- T\. Chen and C\. Guestrin \(2016\)Xgboost: a scalable tree boosting system\.InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining,pp\. 785–794\.Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- T\. Cover and P\. Hart \(1967\)Nearest neighbor pattern classification\.IEEE transactions on information theory13\(1\),pp\. 21–27\.Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- T\. Dettmers, M\. Lewis, Y\. Belkada, and L\. Zettlemoyer \(2022a\)LLM\.int8\(\): 8\-bit matrix multiplication for transformers at scale\.arXiv preprint arXiv:2208\.07339\.Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- T\. Dettmers, M\. Lewis, S\. Shleifer, and L\. Zettlemoyer \(2022b\)8\-bit optimizers via block\-wise quantization\.9th International Conference on Learning Representations, ICLR\.Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. Zettlemoyer \(2023\)Qlora: efficient finetuning of quantized llms\.arXiv preprint arXiv:2305\.14314\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p2.1),[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- P\. Dissanayake and S\. Dutta \(2025\)TabDistill: distilling transformers into neural nets for few\-shot tabular classification\.arXiv preprint arXiv:2511\.05704\.Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- E\. Fix and J\. L\. Hodges Jr \(1951\)Discriminatory analysis\. nonparametric discrimination: small sample performance\.Technical reportCalifornia Univ Berkeley\.Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- E\. Frantar, S\. Ashkboos, T\. Hoefler, and D\. Alistarh \(2022\)GPTQ: accurate post\-training compression for generative pretrained transformers\.arXiv preprint arXiv:2210\.17323\.Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1),[§3\.1](https://arxiv.org/html/2607.27546#S3.SS1.p1.1)\.
- Y\. Fu, L\. Xue, Y\. Huang, A\. Brabete, D\. Ustiugov, Y\. Patel, and L\. Mai \(2024\)\{\\\{serverlessllm\}\\\}:\{\\\{low\-Latency\}\\\}serverless inference for large language models\.In18th USENIX Symposium on Operating Systems Design and Implementation \(OSDI 24\),pp\. 135–153\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p2.1)\.
- A\. Gersho and R\. M\. Gray \(1991\)Vector quantization and signal compression\.Vol\.159,Springer Science & Business Media\.Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p2.1)\.
- A\. Gholami, S\. Kim, Z\. Dong, Z\. Yao, M\. W\. Mahoney, and K\. Keutzer \(2022\)A survey of quantization methods for efficient neural network inference\.InLow\-power computer vision,pp\. 291–326\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p2.1),[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px2.p1.1)\.
- Y\. Gorishniy, A\. Kotelnikov, and A\. Babenko \(2025\)TabM: advancing tabular deep learning with parameter\-efficient ensembling\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=Sd4wYYOhmY)Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- L\. Grinsztajnet al\.\(2025\)TabPFN\-2\.5: advancing the state of the art in tabular foundation models\.External Links:2511\.08667,[Document](https://dx.doi.org/10.48550/arXiv.2511.08667)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1),[§1](https://arxiv.org/html/2607.27546#S1.p3.1),[§1](https://arxiv.org/html/2607.27546#S1.p7.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2),[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px1.p1.1)\.
- S\. Han, H\. Mao, and W\. J\. Dally \(2016\)Deep compression: compressing deep neural networks with pruning, trained quantization and huffman coding\.External Links:1510\.00149,[Link](https://arxiv.org/abs/1510.00149)Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p2.1),[§3\.1](https://arxiv.org/html/2607.27546#S3.SS1.p1.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px2.p1.1)\.
- N\. Herrmann, J\. Stenkamp, B\. Karic, S\. Oehmcke, and F\. Gieseke \(2025\)Boosted trees on a diet: compact models for resource\-constrained devices\.arXiv preprint arXiv:2510\.26557\.Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- G\. Hinton, O\. Vinyals, and J\. Dean \(2015\)Distilling the knowledge in a neural network\.arXiv preprint arXiv:1503\.02531\.Cited by:[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px2.p1.1)\.
- N\. Hollmann, S\. Müller, K\. Eggensperger, and F\. Hutter \(2023\)TabPFN: a transformer that solves small tabular classification problems in a second\.InInternational Conference on Learning Representations 2023,Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- N\. Hollmann, S\. Müller, L\. Purucker, A\. Krishnakumar, M\. Körfer, S\. B\. Hoo, R\. T\. Schirrmeister, and F\. Hutter \(2025\)Accurate predictions on small data with a tabular foundation model\.Nature\.External Links:[Document](https://dx.doi.org/10.1038/s41586-024-08328-6),[Link](https://www.nature.com/articles/s41586-024-08328-6)Cited by:[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- S\. B\. Hoo, S\. Müller, D\. Salinas, and F\. Hutter \(2025\)From tables to time: extending tabpfn\-v2 to time series forecasting\.arXiv preprint arXiv:2501\.02945\.Cited by:[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- E\. J\. Hu, yelong shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px2.p1.1)\.
- I\. Hubara, M\. Courbariaux, D\. Soudry, R\. El\-Yaniv, and Y\. Bengio \(2016\)Binarized neural networks\.InAdvances in Neural Information Processing Systems,D\. Lee, M\. Sugiyama, U\. Luxburg, I\. Guyon, and R\. Garnett \(Eds\.\),Vol\.29,pp\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2016/file/d8330f857a17c53d217014ee776bfd50-Paper.pdf)Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p2.1)\.
- B\. Jacob, S\. Kligys, B\. Chen, M\. Zhu, M\. Tang, A\. Howard, H\. Adam, and D\. Kalenichenko \(2018\)Quantization and training of neural networks for efficient integer\-arithmetic\-only inference\.InProceedings of the IEEE conference on computer vision and pattern recognition,pp\. 2704–2713\.Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p2.1)\.
- B\. P\. Jeyaraman \(2025\)Deployment strategies for llms\.InLarge Language Models Ops for Finance: A Practical Guide to Infrastructure, Implementation, and Innovation,pp\. 103–134\.External Links:ISBN 979\-8\-8688\-1700\-7,[Document](https://dx.doi.org/10.1007/979-8-8688-1700-7%5F4),[Link](https://doi.org/10.1007/979-8-8688-1700-7_4)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p2.1)\.
- J\. Jiang, S\. Liu, H\. Cai, Q\. Zhou, and H\. Ye \(2026\)Representation learning for tabular data: a comprehensive survey\.IEEE Transactions on Pattern Analysis and Machine Intelligence\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- G\. Ke, Q\. Meng, T\. Finley, T\. Wang, W\. Chen, W\. Ma, Q\. Ye, and T\. Liu \(2017\)Lightgbm: a highly efficient gradient boosting decision tree\.Advances in neural information processing systems30\.Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- M\. Li\*, Y\. Lin\*, Z\. Zhang\*, T\. Cai, X\. Li, J\. Guo, E\. Xie, C\. Meng, J\. Zhu, and S\. Han \(2025\)SVDQuant: absorbing outliers by low\-rank components for 4\-bit diffusion models\.InThe Thirteenth International Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- J\. Lin, J\. Tang, H\. Tang, S\. Yang, W\. Chen, W\. Wang, G\. Xiao, X\. Dang, C\. Gan, and S\. Han \(2024\)AWQ: activation\-aware weight quantization for llm compression and acceleration\.InMLSys,Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1),[§3\.1](https://arxiv.org/html/2607.27546#S3.SS1.p1.1)\.
- S\. Liu and H\. Ye \(2025\)TabPFN unleashed: a scalable and effective solution to tabular classification problems\.InForty\-second International Conference on Machine Learning,Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p3.1)\.
- Z\. Liu, C\. Zhao, I\. Fedorov, B\. Soran, D\. Choudhary, R\. Krishnamoorthi, V\. Chandra, Y\. Tian, and T\. Blankevoort \(2025\)SpinQuant: llm quantization with learned rotations\.InThe Thirteenth International Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1)\.
- G\. Menghani \(2023\)Efficient deep learning: a survey on making deep learning models smaller, faster, and better\.ACM Comput\. Surv\.55\(12\)\.External Links:ISSN 0360\-0300,[Link](https://doi.org/10.1145/3578938),[Document](https://dx.doi.org/10.1145/3578938)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p2.1)\.
- A\. C\. Mueller, C\. Curino, and R\. Ramakrishnan \(2025\)MotherNet: fast training and inference via hyper\-network transformers\.InInternational Conference on Learning Representations,Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- S\. Müller, N\. Hollmann, S\. P\. Arango, J\. Grabocka, and F\. Hutter \(2022\)Transformers can do bayesian inference\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=KSugKcbNf9)Cited by:[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- L\. Prokhorenkova, G\. Gusev, A\. Vorobev, A\. V\. Dorogush, and A\. Gulin \(2018\)CatBoost: unbiased boosting with categorical features\.Advances in neural information processing systems31\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p5.1),[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- J\. Qu, D\. Holzmüller, G\. Varoquaux, and M\. L\. Morvan \(2025\)TabICL: a tabular foundation model for in\-context learning on large data\.InForty\-second International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=0VvD1PmNzM)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- J\. Qu, D\. Holzmüller, G\. Varoquaux, and M\. L\. Morvan \(2026\)TabICLv2: a better, faster, scalable, and open tabular foundation model\.External Links:2602\.11139,[Link](https://arxiv.org/abs/2602.11139)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.
- M\. Rastegari, V\. Ordonez, J\. Redmon, and A\. Farhadi \(2016\)XNOR\-net: imagenet classification using binary convolutional neural networks\.InComputer Vision – ECCV 2016,B\. Leibe, J\. Matas, N\. Sebe, and M\. Welling \(Eds\.\),Cham,pp\. 525–542\.External Links:ISBN 978\-3\-319\-46493\-0Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p2.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px2.p1.1)\.
- J\. Seward \(1996\)Bzip2 and Libbzip2\.Note:Publication Title: bzip2 : HomeExternal Links:[Link](https://sourceware.org/bzip2/)Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- Y\. Shi, G\. Ke, Z\. Chen, S\. Zheng, and T\. Liu \(2022\)Quantized training of gradient boosting decision trees\.Advances in neural information processing systems35,pp\. 18822–18833\.Cited by:[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- R\. Shwartz\-Ziv and A\. Armon \(2021\)Tabular data: deep learning is not all you need\.In8th ICML Workshop on Automated Machine Learning \(AutoML\),External Links:[Link](https://openreview.net/forum?id=vdgtepS1pV)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p1.1)\.
- J\. Vanschoren, J\. N\. van Rijn, B\. Bischl, and L\. Torgo \(2013\)OpenML: networked science in machine learning\.SIGKDD Explorations15\(2\),pp\. 49–60\.External Links:[Document](https://dx.doi.org/10.1145/2641190.2641198),[Link](https://doi.org/10.1145/2641190.2641198)Cited by:[§3\.2](https://arxiv.org/html/2607.27546#S3.SS2.p1.1)\.
- G\. Zabërgja, R\. Kamel, A\. Kadra, C\. M\. Frey, and J\. Grabocka \(2026\)End\-to\-end compression for tabular foundation models\.arXiv preprint arXiv:2602\.05649\.Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2),[§2\.3](https://arxiv.org/html/2607.27546#S2.SS3.p1.1)\.
- A\. Zandieh, M\. Daliri, M\. Hadian, and V\. Mirrokni \(2026\)TurboQuant: online vector quantization with near\-optimal distortion rate\.InInternational Conference on Learning Representations \(ICLR\),Note:Preprint available at arXiv:2504\.19874Cited by:[§2\.2](https://arxiv.org/html/2607.27546#S2.SS2.p3.1),[§5](https://arxiv.org/html/2607.27546#S5.SS0.SSS0.Px1.p1.1)\.
- Y\. Zeng, W\. Kang, and A\. C\. Mueller \(2024\)TabFlex: scaling tabular learning to millions with linear attention\.InNeurIPS 2024 Third Table Representation Learning Workshop,External Links:[Link](https://openreview.net/forum?id=f8aganC0tN)Cited by:[§1](https://arxiv.org/html/2607.27546#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.27546#S2.SS1.p1.2)\.

## Appendix AAdditional Details

### A\.1Datasets

All datasets were evaluated by standardised train:test split as provided by Open ML dataset tasks\. Details of the datasets used are provided below:

### A\.2Classical Comparisons

#### Baseline Methods

Baseline models were run for comparison using fixed default configurations across multiple independent datasets\. The results provided are the mean across all datasets for each model\. Chosen parameters are included for reference:

```
XGBoost
        n_estimators: 100
        max_depth: 6
        learning_rate:0.1
        random_state: 42
        n_jobs: -1
        eval_metric: ’logloss’
        objective: ’binary:logistic’ (binary) or ’multi:softmax’ (multi-class)
        use_label_encoder: False

    CatBoost
        iterations: 200
        depth: 6
        learning_rate: 0.1
        random_state: 42
        verbose: False
        thread_counts: -1
        loss_function: ’Logloss’ (binary) or ’Multiclass’ (multi-class)

    LightGBM
        n_estimators: 100
        max_depth: 6
        learning_rate:0.1
        random_state: 42
        n_jobs: -1
        verbose: -1
        force_col_wise: True

    RandomForest
        n_estimators: 100
        max_depth: 10
        random_state: 42
        n_jobs: -1
        verbose: 0

    KNN
        n_neighbors: 5
        n_jobs: -1
```

#### Tuning Methods

RandomisedSearchCVwas used with stratified 3\-fold cross validation \(20 random hyperparameter configurations per model\) using the training dataset and reporting the optimized model performance on the test data\. Multiple parameter grid search spaces were used covering regularization, tree depth, learning rates, and other model\-specific parameters\. Details of the parameter search grids are included below:

```
XGBoost (9 parameters)
        max_depth: [3, 5, 7, 10]
        learning_rate: [0.01, 0.05, 0.1, 0.2]
        n_estimators: [50, 100, 200, 300]
        min_child_weight: [1, 3, 5]
        subsample: [0.6, 0.8, 1.0]
        colsample_bytree: [0.6, 0.8, 1.0]
        gamma: [0. 0.1, 0.2]
        reg_alpha: [0, 0.1, 0.5, 1.0]
        reg_lambda: [0.5, 1.0, 2.0]

    CatBoost (7 parameters)
        iterations: [50, 100, 200, 300]
        depth: [4, 6, 8, 10]
        learning_rate: [0.01, 0.05, 0.1, 0.2]
        l2_leaf_reg: [1, 3, 5, 7]
        border_count: [32, 64, 128]
        bagging_temperature: [0, 0.5, 1.0]
        random_strength: [0.5, 1.0, 2.0]

    LightGBM (9 parameters)
        num_leaves: [15, 31, 63, 127]
        max_depth: [3, 5, 7, 10, -1]
        learning_rate: [0.01, 0.05, 0.1, 0.2]
        n_estimators: [50, 100, 200, 300]
        min_child_samples: [10, 20, 30, None]
        subsample: [0.6, 0.8, 1.0]
        colsample_bytree: [0.6, 0.8, 1.0]
        reg_alpha: [0, 0.1, 0.5, 1.0]
        reg_lambda: [0, 0.1, 0.5, 1.0]

    RandomForest (7 parameters)
        n_estimators: [50, 100, 200, 300]
        max_depth: [5, 10, 20, 30, None]
        min_samples_split: [2, 5, 10]
        min_samples_leaf: [1, 2, 4]
        max_features: [’sqrt’, ’log2’, 0.5, 0.7]
        bootstrap: [True, False]
        criterion: [’gini’, ’entropy’]

    KNN (5 parameters)
        n_neighbours: [3, 5, 7, 9, 11, 15, 21]
        weights: [’uniform’, ’distance’]
        metric: [’euclidean’, ’manhattan’, ’minkowski’]
        p: [1, 2, 3]
        algorithm: [’ball_tree’, ’kd_tree’, ’brute’]
```

## Appendix BHardware Details

Experiments were conducted on an AWS SageMakerml\.g4dn\.12xlargeinstance \(4×\\timesNVIDIA T4 GPUs with 16 GB VRAM each, 48 vCPUs and 192 GB of shared RAM\)\. Four worker processes were run in parallel; each worker was pinned to one GPU and a dedicated set of 12 vCPUs to prevent inter\-worker contention\. RAM was shared across workers rather than strictly partitioned, as physical memory cannot be isolated in the same way as compute resources\. Per\-process peak resident set size \(RSS\) during inference is reported in Table[4](https://arxiv.org/html/2607.27546#A3.T4)\.

## Appendix CAblations

### C\.1K\-means N\-bit ablation study

Table 2:K\-means N\-bit ablation studyNote: Results computed on 5 representative datasets and may differ slightly from 30\-dataset averages in Table 1 due to dataset sampling\.

4\-bit quantization achieves the optimal trade\-off between compression and performance, recovering over 99% of baseline accuracy at 7\.64× compression\. 2\-bit quantization offers a viable alternative at 14\-15× compression with∼\\sim90% accuracy recovery\. 1\-bit quantization is impractical for production use due to severe accuracy degradation \(∼\\sim40%\)\.

### C\.2Performance stability across datasets

Table 3:Performance stability study across 30 classification datasetsNote: Results are averaged across 30 datasets\. Standard deviation represents performance variation across datasets\.

Standard deviation across datasets remains consistent between quantized and baseline variants, indicating that quantization does not introduce additional performance variance across different data distributions\.

### C\.3Inference memory and speeds

Table 4:Inference memory and speed averaged across 30 OpenML datasets\.\*Inference Mem = Peak RSS \- Baseline RSS \(process overhead removed\)\.

Table[4](https://arxiv.org/html/2607.27546#A3.T4)presents an analysis of inference efficiency\. Our primary contribution focuses on weight compression to address deployment bottlenecks \(model distribution and storage\), achieving up to 7× compression\. As an additional analysis, we measure inference memory, which remains largely unchanged \(<2<2% variation,∼\\sim4\.7 GB\), as expected since memory usage is dominated by activations\. This observation motivates future work on activation quantization\. Inference speed remains stable \(±10%\), indicating no performance degradation from weight quantization\.