Zer0Fit adds local prediction tools to AI agents
Zer0Fit runs Google's TabFM and TimesFM models in one and connects them to a through MCP. Open WebUI, Claude Code, and Codex can use the setup for forecasting, , and regression without first training or tuning a separate model. Reported tests reached 94.7% accuracy on the Iris dataset and an R-squared score of 0.87 on a .
Running both models requires about 16GB of VRAM on an Nvidia , and the current PyTorch setup supports CUDA only. To save memory, an unused model is automatically unloaded after five minutes. CSV files work now, while XLS, XLSX, JSON, and JSONL support is planned.
The installer detects the computer architecture automatically, and the software was tested on DGX Spark, RTX 3090, and H100 hardware.
Key points
- MCP connects TabFM and TimesFM to .
- The models handle forecasting, , and regression without separate training.
- The reported results were 94.7% Iris accuracy and 0.87 R-squared for regression.
- Both models need about 16GB of VRAM and an Nvidia with CUDA support.
- CSV works now; XLS, XLSX, JSON, and JSONL support is planned.