Google Research's TabFM, a zero-shot foundation model for tabular data, claims to outperform tuned gradient-boosted trees without any training. An independent evaluation was conducted to verify these assertions, utilizing three machines. The results confirmed that the model adheres to the scikit-learn contract, demonstrating bit-exact deterministic predictions. Initial tests showed that TabFM could learn from context effectively, achieving high accuracy on known-answer tasks. This evaluation underscores the need for rigorous independent reproduction in confirming the efficacy of new models.
Tabular DataFoundation ModelsIndependent Evaluation+1
Tabular DataFoundation ModelsIndependent EvaluationMachine Learning
Independent Evaluation of Google's TabFM Foundation Model Yields Insights
Hacker News· yashraj pandey· Tuesday, July 7, 2026
More Articles From This Day
Generative AILLM+2
Generative AILLMAgentsNLP
Tencent Launches Hy3, a 295B-Parameter Mixture-of-Experts Model for Agentic Workflows
Tencent has unveiled Hy3, a 295 billion-parameter Mixture-of-Experts model designed for reasoning and agentic workflows, positioning itself as a strong competitor to trillion-scale models. The model features 21 billion active parameters with 192 experts and offers a 256K context window aimed at long-horizon tasks such as coding, document processing, and financial analysis. Hy3 is built to mitigate hallucination issues by providing grounded responses and is available for commercial use under the Apache 2.0 license. Users can access a free API for two weeks, enhancing its affordability for various applications.