TabPFN-3.5 ships: tabular predictions that beat XGBoost, no training
#1 on TabArena, up to 1M rows, 0.17s inference — drop it in with pip. Plus: Claude Code adds fast Remote mode and DeepSeek Flash Latest hits OpenRouter.
Copy markdown
The pitch: a foundation model for your spreadsheets
TabPFN-3.5 is a pretrained model you point at structured data — no training loop, no hyperparameter search. It ranks #1 on TabArena and BeyondArena for real-world prediction, posts a 99% win rate over classic ML like gradient boosting, and handles up to 1M rows natively. Aim it at churn, fraud, or lead-scoring and get a strong baseline from a single fit().
The speed: 0.17s per 1,000 rows
TabPFN-3.5-Fast returns predictions in 0.17s per 1k rows (0.5s for standard 3.5) — roughly 20x faster than the previous generation. That's quick enough to sit in a request path, not just an overnight batch job.
How to run it: pip install tabpfn
Same scikit-learn fit/predict API, with weights on Hugging Face. The license catch to check before shipping: the newest weights are non-commercial, so prototype freely but route commercial use through the hosted API (the older 2.x line carries an Apache-flavored license).
Elsewhere: Claude Code 2.1.271 adds fast Remote mode
The new build drops a /fast command into Claude Code Remote sessions, renders markdown artifacts as styled document pages, and adds per-command allowed_domains for Bash and PowerShell when running sandboxed in auto mode.
Elsewhere: DeepSeek Flash Latest lands on OpenRouter
DeepSeek added a deepseek-flash-latest alias on OpenRouter — an auto-updating pointer to its newest Flash model, so you can track the latest release without editing model IDs in your code.