A grad student put Google's new 'no-training' ML in a chat box

The "no-training" line keeps moving. Google put it in front of time-series data with TimesFM, a time-series foundation model, then extended the same trick to spreadsheet-style "tabular" data on June 30 with TabFM, a hybrid-attention foundation model by Weihao Kong and Abhimanyu Das. Four weeks later, the models are already in a chat box.

Zer0Fit, a weekend project from an AI grad student using the handle porespellar, wraps TabFM v1.0.0 and TimesFM 2.5 as a single Model Context Protocol (MCP) server, an open standard that lets a chat client treat an external model as a tool. The whole thing ships in one Docker container. A user drops a CSV into Open WebUI, asks the model to classify a column or forecast a series, and gets an answer without writing training code, picking a learning rate, or splitting a train set.

That is the actual shift, and it is small enough to miss if you read the wrapper as a one-off hobby project. A research-grade foundation model on one side, an MCP harness on the other, and the build, train, and tune barrier for tabular machine learning collapses between them. The wrapper does not invent the models and it does not need to. Its job is plumbing.

TabFM is a foundation model in the same sense as a large language model: a single pre-trained network that handles many tasks. Where an LLM is pre-trained on text, TabFM is pre-trained on synthetic and public tabular datasets, then asked to classify rows or predict numeric columns in CSVs it has never seen. "Zero-shot" here means the model performs the task on a reader's data without being retrained on that data. The PyTorch weights are mirrored on Hugging Face and the reference code is open under an Apache-style license, so anyone with the right GPU can load them.

Zer0Fit's contribution is making that load step boring. The install script auto-detects whether it is running on an NVIDIA DGX Spark (ARM plus CUDA 13) or a 3090-class card on AMD64 (CUDA 12.6). Models are loaded and unloaded on demand with a five-minute TTL, which matters because TabFM and TimesFM do not both fit comfortably on a single 16GB card. CSV is the only format the server accepts today; the developer lists XLS, XLSX, JSON, and JSONL as forthcoming on the project README. Open WebUI is the primary tested client; integrations for Claude Code and Codex CLI are in the repo but less exercised.

The headline number on the announcement post is 94.7% accuracy on the Iris flower dataset and an R² of 0.91 on California Housing, both measured by the developer against traditionally tuned baselines. Those are encouraging, and they are also the wrong comparison for a working data scientist. Iris has 150 rows, four columns, and three classes; California Housing is a 1990s teaching regression. Independent evaluation is still thin. MarkTechPost's same-week explainer walks the architecture but does not run the model; Yashraj Pandey's hands-on is the only third-party attempt so far. Replication on a messier dataset is the next thing the community owes the project.

The honest hardware floor is also the part most coverage will skip. The underlying Google checkpoints are PyTorch, which means CUDA, Nvidia's GPU compute layer, which excludes Macs and AMD or Intel GPU stacks. Sixteen gigabytes of VRAM (GPU memory) is the working minimum. That excludes a large slice of the developer audience that has been told for two years that local ML is the next normal. A reader with an M-series MacBook, a Steam Deck, or a corporate laptop is not in this user base. The repo carries a matching disclaimer: "use at your own risk," and it bans financial, medical, legal, safety-critical, and employment decisions. The wrapper is research and education only.

None of that takes the shift back. Google has now put two zero-shot foundation models in the same family, both with open weights, and a single Docker pull is enough to wire them into a chat client on a home GPU. The pattern is what to watch: every time a research lab drops a foundation model, the time from release to "callable from a chat window" is measured in weeks, and the audience for tabular machine learning is whatever builds the next thin wrapper. The next honest test is a CSV nobody in this thread has seen yet, run on a card somebody already owns.