Today we're opening early access to the hosted PredictLM API — the same open-weight tabular foundation models you can already pip install, now behind a single HTTPS endpoint.
What it does
One call, no training step:
POST /v1/predict
{
"X_train": [[5.1, 3.5, 1.4], ...],
"y_train": [0, 0, 1, ...],
"X_query": [[6.2, 2.9, 4.3]]
}You pass a small training table and the rows you want predicted. The model does in-context learning in a single forward pass — no fit, no hyperparameter sweep, no model selection. Float targets route to regression, integer targets to classification, and classification returns calibrated per-class probabilities, not just labels.
Under the hood it's predictlm-base-26m, the Apache-2.0 checkpoint already on Hugging Face. When a stronger model passes our public-benchmark and latency gates, it swaps in behind the same contract — the endpoint you integrate today is the endpoint that gets better underneath you.
Why an API when the weights are free?
The weights stay free — that doesn't change. The API is for the cases where you don't want to own inference: spreadsheets and internal tools calling over HTTP, agents using prediction as a tool call, pipelines that want a dependency-free integration, and teams that want the current-best model without tracking our releases.
Early access, deliberately small
We're onboarding by hand. You tell us what you're building, we send you a key — usually within a day. No self-serve signup yet, no credit card. The first users shape the roadmap: rate limits, batch endpoints, distribution outputs, and what we build next (imputation and anomaly scoring fall out of the next-generation architecture naturally — more on that soon).
If you'd rather run it yourself, everything ships as before: models on Hugging Face, pip install predictlm, MCP server on GitHub.