Zer0Fit wraps Google's TabFM and TimesFM as a local MCP server

A developer has created Zer0Fit, an open-source implementation that serves Google's TabFM and TimesFM foundation models as a Model Context Protocol (MCP) server for zero-shot machine learning tasks. The project allows users to perform forecasts, classifications, and regressions locally without building or training custom models.

• Combines Google's TabFM (tabular data) and TimesFM (time series) transformer models into a single Docker container.
• Achieved 94.7% accuracy on the Iris dataset and an R2 of 0.87 on California Housing regression in zero-shot testing.
• Requires approximately 16GB of VRAM and supports dynamic model loading with a 5-minute TTL to manage memory.
• Compatible with Open WebUI, Claude Code, and Codex, running on CUDA-enabled Nvidia hardware such as DGX Spark, 3090, and H100.

This tool enables users to connect local LLMs to ML foundation models for tasks that previously required extensive hyperparameter tuning and model training.