01
Context
Pass a small set of labeled rows from your tabular dataset — typically 5 to 500. Mixed numeric and categorical features are handled natively.
Zero One Research/PredictLM
The agent-native tabular foundation model · Built in the EU
Calibrated predictions for tabular data, inside the agent stack.
PredictLM is the smallest open-weight tabular foundation model with calibrated uncertainty out of the box. 13M parameters, Apache-2.0, runs on a laptop. Ships as a Python sklearn-style class and a Claude / Cursor / Continue MCP tool — give your LLM agent a real numerical co-processor instead of letting it hallucinate predictions. EU-based and GDPR-native — deployable for healthcare, finance, and public-sector workloads that can't ship data offshore.
- Parameters
- 13M / 26M
- License
- Apache-2.0
- Uncertainty
- Calibrated
- MCP tool
- Day-1 ready
How it works
Three steps. One forward pass.
No training loop. No hyperparameter search. PredictLM treats your labeled rows as context and conditions on them the way a language model conditions on a prompt.
- 02
In-context inference
The transformer attends across features and rows simultaneously, building an implicit predictor for your dataset on the fly.
- 03
Predict
Predictions for new rows come out immediately, with calibrated uncertainty — full predictive distributions for regression, class probabilities for classification.
The model family
Two checkpoints. Same interface.
Base prioritizes raw accuracy. Mini is tuned to run on a laptop CPU — and stays statistically tied with Base on classification.
PredictLM
Base 26M
- Parameters26M
- Checkpoint105 MB
- LicenseApache-2.0
Evaluation on 25 OpenML datasets
Regression
R² 0.589 mean
R² 0.755 median
Classification
acc 0.685 mean
acc 0.799 median
Competitive with XGBoost on regression — point estimate ahead but CI overlap is within sampling noise.
PredictLM
Mini 13M
- Parameters13M
- Checkpoint~54 MB
- LicenseApache-2.0
Versus Base — same 25-dataset benchmark
Classification
Tied delta −0.001
CI [−0.027, +0.029]
Regression
−4 pp R²
CI [−6.5, −1.5]
~95% of Base's quality at half the size. Runs comfortably on CPU.
Quickstart
From pip install to predictions in under a minute.
Standard transformers loading. No custom training loop, no per-dataset config. Works on CPU for Mini.
predict.pypython
# pip install predictlm
from predictlm import PredictLM
# One model object — the partner ckpt is downloaded on first .predict()
# and the package runs the published Duo + TTT ensemble under the hood.
model = PredictLM.from_pretrained("zerooneresearch/predictlm-mini-13m")
# Your data: ~50 labeled rows + new rows to predict
# X_train, y_train, X_test as numpy arrays or pandas
# Just .fit().predict() — returns the 0.751 cls / 0.609 reg result
# on the locked 25-dataset OpenML benchmark.
preds = model.fit(X_train, y_train).predict(X_test)
# Need single-model latency? Pass auto_duo=False at load:
# model = PredictLM.from_pretrained(..., auto_duo=False) # ~0.673 clsUse cases
Built for the spots where traditional ML breaks down.
Tabular ML without training
Drop-in regression or classification on a brand-new dataset in seconds. Skip the hyperparameter search, skip the cross-validation harness — just call predict.
Small-data settings
Sweet spot is roughly 10–500 labeled rows — the regime where boosted trees often overfit and deep models simply don't converge.
Calibrated uncertainty
Our calibrated regression head returns a full predictive distribution over the target — not just a point estimate. Useful for risk-aware downstream decisions.
Evaluation
Recipe comparison
Mean and median performance across the locked 25-dataset OpenML benchmark (10 regression + 10 classification, fair-set filter n_features ≤ 128, seed = 42). The PredictLM weights are the same in every row below — only the inference recipe changes.
| Model | Params | Reg. R² (mean) | Reg. R² (median) | Cls. acc (mean) | Cls. acc (median) |
|---|---|---|---|---|---|
| PredictLM-Mini (zero-tune) | 13M | 0.536 | 0.657 | 0.673 | 0.747 |
| PredictLM-Base (zero-tune) | 26M | 0.589 | 0.755 | 0.685 | 0.799 |
| Mini + TTT | 13M | 0.595 | 0.661 | 0.742 | 0.733 |
| Base + TTT | 26M | 0.608 | 0.671 | 0.748 | 0.756 |
| Duo + TTT (w=0.40) | 13M + 26M | 0.609 | 0.673 | 0.751 | 0.763 |
| XGBoost (baseline) | — | 0.561 | 0.744 | 0.679 | 0.793 |
Bold entries mark the per-column best. The bottom three rows use the same Mini and Base weights as the top two — only the inference recipe differs. TTT applies ~15 inner Adam steps of self-supervised fine-tuning on the user's in-context examples per task.
Citation
Citing PredictLM.
If PredictLM helps your work, a citation is appreciated.
predictlm.bibbibtex
@misc{predictlm2026,
author = {Svoboda, M. and collaborators},
title = {PredictLM: Tabular Foundation Models for In-Context Prediction},
year = {2026},
url = {https://huggingface.co/zerooneresearch}
}