NVIDIA发布Kumo Tabular：单前向推理的开源表格基础模型
NVIDIA Releases Kumo Tabular: Open Tabular Foundation Models That Predict New Rows in a Single Forward Pass
表格数据处理是工业界刚需，Kumo Tabular以零训练成本实现高性能预测，且开放商用许可，对落地应用极具价值，建议关注其实际部署效果。
NVIDIA has released Kumo Tabular, a new family of tabular foundation models (TFMs) for classification and regression. If you have followed TabPFN or TabICL, the setup will look familiar. The model takes labeled rows as context and predicts new rows in one forward pass. There is no training, no hyperparameter tuning, and no feature engineering.
NVIDIA 发布了 Kumo Tabular，这是一个用于分类和回归的新型表格基础模型（TFM）家族。如果你关注过 TabPFN 或 TabICL，其设置会看起来很熟悉。该模型将带标签的行作为上下文，并在一次前向传播中预测新行。无需训练、超参数调优或特征工程。
Kumo Tabular comes in Small, Medium, and Large versions, spanning about 28M to 215M parameters. It runs through NVIDIA’s open-source structured-data-models (SDM) library.
Kumo Tabular 提供 Small、Medium 和 Large 版本，参数量范围约为 28M 至 215M。它通过 NVIDIA 的开源结构化数据模型（SDM）库运行。
Is it deployable? Yes. Weights ship under the OpenMDW-1.1 license, which permits commercial use. The SDM code is Apache-2.0, and it needs Python 3.11+ and PyTorch 2.7+, with examples targeting a CUDA GPU.
可以部署吗？可以。权重采用 OpenMDW-1.1 许可证发布，允许商业使用。SDM 代码采用 Apache-2.0 许可证，需要 Python 3.11+ 和 PyTorch 2.7+，示例针对 CUDA GPU。
What the SDM Library Adds
SDM 库带来的增强
SDM is a GPU-native library for structured-data foundation models and preprocessing. Besides Kumo Tabular, it ships TabICLv2, Google’s TabFM, and KumoRelational for multi-table data. All models share one in-context learning interface built on a TableTensor container. The library also handles preprocessing, ensembling, and many-class prediction.
SDM 是一个面向结构化数据基础模型及预处理的 GPU 原生库。除了 Kumo Tabular，它还包含 TabICLv2、Google 的 TabFM 以及用于多表数据的 KumoRelational。所有模型都共享一个基于 TableTensor 容器的上下文学习接口。该库还处理预处理、集成学习和多类别预测。
How Kumo Tabular Works
Kumo Tabular 的工作原理
Kumo Tabular is a Transformer built around the structure of a table. It uses column, row, and in-context attention, as introduced in TabICL and TabPFN. The pipeline has 3 stages:
Kumo Tabular 是一个围绕表格结构构建的 Transformer。它使用了在 TabICL 和 TabPFN 中引入的列注意力、行注意力和上下文注意力。该流程分为三个阶段：
- Cell embedding: Numerical and categorical values pass through learned Fourier features, with separate weights per type. Missing values need no imputation.
- Row embedding: Column attention uses induced self-attention, so cost grows linearly with rows. Row attention, with rotary positions, learns feature interactions. 4 learnable [CLS] tokens compress each row.
- In-context learning: A final Transformer runs over row embeddings. Context rows attend to each other, while query rows attend only to context rows.
- 单元格嵌入：数值型和分类型值通过学习的傅里叶特征传递，每种类型拥有独立的权重。缺失值无需插补。
- 行嵌入：列注意力使用诱导自注意力，因此计算成本随行数线性增长。带有旋转位置编码的行注意力学习特征交互。4 个可学习的 [CLS] token 压缩每一行。
- 上下文学习：最终的 Transformer 在行嵌入上运行。上下文行之间相互注意，而查询行仅注意上下文行。
Because the context never sees the queries, its keys and values are computed once and reused. The head outputs class probabilities, or 999 quantiles for regression. That gives a point prediction plus an uncertainty estimate.
由于上下文无法看到查询，其键（keys）和值（values）只计算一次并重复使用。输出头输出类别概率，或者回归任务的 999 个分位数。这提供了点预测以及不确定性估计。
One more detail matters at scale. Softmax attention spreads thin as the number of keys grows. Kumo Tabular scales each query by a temperature that grows with the log of the key count. The coefficient is learned per attention head, so attention stays sharp on larger tables.
规模扩展时还有一个细节很重要。随着键数量的增加，Softmax 注意力会变得分散。Kumo Tabular 根据键数量的对数增长来缩放每个查询的温度系数。该系数按注意力头学习，因此在更大的表格上注意力仍能保持尖锐。
Trained Only on Artificial Tables
仅在人工表格上训练
Kumo Tabular is pretrained entirely on synthetic tables sampled from Structural Causal Models (SCMs). A random causal graph links hidden variables through linear maps, small neural networks, trees, or Gaussian processes. The generator also injects messy, real-world patterns: missing values, high-cardinality categories, heavy-tailed targets, and conflicting duplicate rows.
Kumo Tabular 完全在从结构因果模型（SCMs）采样的合成表格上进行预训练。一个随机的因果图通过线性映射、小型神经网络、树或高斯过程将隐藏变量连接起来。生成器还注入了混乱的真实世界模式：缺失值、高基数类别、重尾目标以及冲突的重复行。
Training ran in 3 stages, similar to TabICLv2. Context grew from 1,024 rows to 60,000 rows, with up to 100 columns. Small, Medium, and Large saw about 35M, 71M, and 137M artificial tables. Classification and regression are trained as separate models. NVIDIA says the training recipe and data generators will be released soon.
训练分为三个阶段，类似于 TabICLv2。上下文从 1,024 行增长到 60,000 行，最多包含 100 列。Small、Medium 和 Large 版本分别处理了约 35M、71M 和 137M 个合成表格。分类和回归作为独立的模型进行训练。NVIDIA 表示，训练配方和数据生成器即将发布。
Benchmarks
基准测试
With default settings, Kumo Tabular ranks first overall on TabArena with an Elo of 1950. NVIDIA team reports it runs 17x faster than LimiX-2 on a single RTX 6000 Pro. All 3 sizes sit on the accuracy and inference-time Pareto front.
在默认设置下，Kumo Tabular 在 TabArena 上以 1950 的 Elo 评分位居整体第一。NVIDIA 团队报告称，它在单张 RTX 6000 Pro 上的运行速度比 LimiX-2 快 17 倍。所有三个尺寸均位于准确率与推理时间的帕累托前沿。
- BeyondArena: First place, with an Elo of 1418 and an Improvability score of 7.78%.
- TALENT: Top overall ranking, with average ranks of 6.67 (accuracy), 3.98 (log-loss), and 4.22 (RMSE).
- ScoringBench: Large and Medium rank first and second on average rank.
- BeyondArena：排名第一，Elo 为 1418，可改进性得分为 7.78%。
- TALENT：总体排名最高，平均排名分别为 6.67（准确率）、3.98（对数损失）和 4.22（RMSE）。
- ScoringBench：Large 和 Medium 版本在平均排名中分列第一和第二。
Kumo Tabular vs Its Closest Competitors
Kumo Tabular 与其最接近的竞争者对比
| Feature | Kumo Tabular | TabICLv2 | TabPFN-3 | LimiX-2 | TabFM |
|---|---|---|---|---|---|
| Developer | NVIDIA | Inria SODA | Prior Labs | Stable AI | Google Research |
| Parameters | ~28M to 215M (3 sizes) | 27.55M (cls), 28.54M (reg) | Not listed in docs | 400M | ~1.64B |
| Tasks | Classification, regression | Classification, regression | Classification, regression | Classification, regression, imputation | Classification, regression |
| Native classes per pass | 10 (ECOC for more) | 10 (hierarchical for more) | 160 | Not specified | 10 (hard limit) |
| Weights license | OpenMDW-1.1 | BSD-3-Clause | TABPFN-3 License v1.0 | StableAI LimiX Non-Commercial | TabFM Non-Commercial v1.0 |
| Commercial use of weights | Yes | Yes | Paid license required | No | No |
| Runs in NVIDIA SDM | Yes | Yes | No | No | Yes |
| 特性 | Kumo Tabular | TabICLv2 | TabPFN-3 | LimiX-2 | TabFM |
|---|---|---|---|---|---|
| 开发者 | NVIDIA | Inria SODA | Prior Labs | Stable AI | Google Research |
| 参数量 | ~28M 至 215M（3 种尺寸） | 27.55M（分类），28.54M（回归） | 文档未列出 | 400M | ~1.64B |
| 任务 | 分类、回归 | 分类、回归 | 分类、回归 | 分类、回归、插补 | 分类、回归 |
| 每轮原生类别数 | 10（ECOC 支持更多） | 10（分层支持更多） | 160 | 未指定 | 10（硬性限制） |
| 权重许可证 | OpenMDW-1.1 | BSD-3-Clause | TABPFN-3 License v1.0 | StableAI LimiX 非商业 | TabFM 非商业 v1.0 |
| 权重的商业使用 | 是 | 是 | 需付费许可 | 否 | 否 |
| 是否在 NVIDIA SDM 中运行 | 是 | 是 | 否 | 否 | 是 |
Sources: NVIDIA blog, SDM model docs, Prior Labs docs, LimiX GitHub, TabICLv2 paper. Checked September 30, 2026.
来源：NVIDIA 博客、SDM 模型文档、Prior Labs 文档、LimiX GitHub、TabICLv2 论文。核查日期为 2026 年 9 月 30 日。
The license row is the real differentiator. TabPFN-3, LimiX-2, and TabFM weights carry non-commercial terms. Kumo Tabular and TabICLv2 are the permissive options, and Kumo Tabular leads the benchmarks NVIDIA reports.
许可证行才是真正的区别因素。TabPFN-3、LimiX-2 和 TabFM 的权重带有非商业条款。Kumo Tabular 和 TabICLv2 是宽松许可选项，且 Kumo Tabular 在 NVIDIA 报告的基准测试中领先。
Getting Started
入门指南
Install the library and pass a DataFrame through TableTensor, following the model card:
安装库并通过 TableTensor 传递 DataFrame，遵循模型卡片说明：
# pip install structured-data-models
from sklearn.datasets import load_breast_cancer
import sdm
df = load_breast_cancer(as_frame=True).frame
table = sdm.TableTensor.from_pandas(
df=df,
stypes=sdm.infer_stypes(df, overrides={"target": "categorical"}),
device="cuda",
)
model = sdm.models.KumoTabular(task="classification", device="cuda")
probs = model(
x_context=table[:300].drop_columns("target"),
y_context=table[:300, "target"],
x_query=table[300:].drop_columns("target"),
num_estimators=8,
)The size argument accepts "small", "medium", or "large", and defaults to large.
size 参数接受 "small"、"medium" 或 "large"，默认值为 large。
Key Takeaways
要点
- NVIDIA’s Kumo Tabular predicts new table rows in 1 forward pass, with no training.
- 3 sizes span about 28M to 215M parameters, pretrained only on synthetic tables.
- NVIDIA reports first place on TabArena (Elo 1950), BeyondArena, TALENT, and ScoringBench.
- OpenMDW-1.1 weights allow commercial use, unlike TabPFN-3, LimiX-2, and TabFM.
- It runs through NVIDIA’s GPU-native SDM library alongside TabICLv2, TabFM, and KumoRelational.
- NVIDIA 的 Kumo Tabular 可在一次前向传播中预测新的表格行，无需训练。
- 三种尺寸涵盖约 28M 至 215M 参数，仅在合成表格上进行预训练。
- NVIDIA 报告在 TabArena（Elo 1950）、BeyondArena、TALENT 和 ScoringBench 上均排名第一。
- OpenMDW-1.1 权重允许商业使用，而 TabPFN-3、LimiX-2 和 TabFM 则不允许。
- 它通过 NVIDIA 的原生 GPU SDM 库运行，与 TabICLv2、TabFM 和 KumoRelational 并列。
Check out the Model on HF, GitHub Repo and Technical details. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看 HF 上的模型、GitHub 仓库和技术细节。所有功劳归于该项目的研究者。此外，欢迎在 Twitter 上关注我们，并别忘了加入我们有 15 万+成员的 ML SubReddit 以及订阅我们的新闻通讯。等等！你在 Telegram 上吗？现在你也可以在 Telegram 上加入我们。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力