Foundation model for tables skips training

Synthefy released Nori V1, a foundation model with open weights aimed at predictions on tabular data, and its main trick is that it requires no training at all. Instead of fitting a new model for each dataset, users pass the test rows together with a marked context (X_train, y_train) and obtain immediate predictions. The code and weights are released under the Apache 2.0 license, free for commercial use.

The interface resembles retrieval more than classic training: `model.predict(X_test, X_train, y_train)` returns predictions without a separate fitting step. The model treats the training examples as context during inference, similar to how language models handle few-shot prompting, but applied to tabular data. Execution runs on a single GPU with automatic fallback to CPU if no GPU is available.

The package is available on pip as `synthefy-nori`. The weights are downloaded once and cached locally, so subsequent runs do not require another download. The Apache 2.0 license permits embedding the model in commercial products without asking anyone or paying royalties, a notable difference from “open” models that restrict commercial use.

The announcement claims “Replaces XGBoost”, but the release does not publish benchmarks against current versions of XGBoost, LightGBM or CatBoost on standard metrics. A foundation model for tables is an interesting direction, especially for tasks with no time or data for dedicated training, yet without numbers on accuracy, inference time and memory consumption compared to established baselines, the claim remains to be validated in practice.

No details have been released about the model architecture, weight size, original training data, or known limitations such as maximum column count, supported data types, or behavior with unseen categories. It is also unclear how the model handles information leakage when the context includes the entire training set. Until independent evaluations appear, Nori V1 remains a curious proof of concept rather than a proven replacement for the boosting stacks the industry has relied on for years.