TabPFN 3.5 Beats Boosted Trees: When to Use It on Your Data

For decades the best way to predict a number or a category from a spreadsheet was to train a gradient-boosted tree model on that spreadsheet. On September 14, 2026 Prior Labs published the technical report for TabPFN-3.5, a model that has already been trained and simply reads your table. The report ranks it first on all seven benchmarks it lists. The numbers are the authors' own, and the licence draws a line between trying it and shipping it.

This post is for two readers: the analyst or founder with a prediction problem the size of a spreadsheet, and the builder wiring a tabular model into an analytics agent. It explains what the model is, prints the benchmark table as the report gives it, and turns the report's own caveats into a decision. Every figure is from the report. We have run no benchmark of our own.

Ordinary machine learning on a table works in two steps. You train a model on your rows, then use it to predict new rows. The training step is where the skill and the time go: choosing the model, tuning it, checking it does not overfit. A tabular foundation model skips that step. It was trained once, by its makers, on a large synthetic collection of generated tables, and learned a procedure for reading a table and predicting a missing column. When you use it, the rows you already know are the context and the rows you want predicted are the question, the way a language model reads a prompt. There is no fitting to your data in the usual sense.

The report states the practical consequences. TabPFN-3.5 accepts training sets of up to one million rows and, in its recommended limits, six thousand columns. It handles classification with many classes, returns a full predictive distribution rather than a point estimate, and supports cached inference in which the known rows are processed once and new rows are scored, in the report's words, "in well under a second". Its parameter count is 220 million, up from 53 million for classification and 58 million for regression in the previous release.

The report's headline table lists seven public benchmarks, the rank of the best TabPFN-3.5 family member on each, and the share of dataset splits it won against the strongest other tabular foundation model and the strongest conventional model. Note the word "family". On four of the seven the ranked entry is the Thinking variant, which is available only through the paid API, not the open weights.

The lowest two bars are the ones to read. BeyondArena is the benchmark built to test tables that are not clean and independent: grouped and temporal splits, tiny and huge tables, text and high-cardinality columns. RelArena is relational data across linked tables. Those are the shapes most business data takes, and they are where the margin is thinnest.

The report is unusually direct about the previous generation's weaknesses. Citing an earlier study, it says earlier tabular foundation models excelled on clean data but fell behind tuned conventional models on grouped or temporal splits, on tables with many rows or many columns, and on tables with text or high-cardinality columns. Its own framing of the new release is that on those regimes TabPFN-3.5 is "now competitive with heavily tuned gradient-boosted trees and MLPs". Competitive is not dominant, and the BeyondArena section says so in detail.

On the full BeyondArena set, the report says tuned and ensembled MLPs "retain the highest performance on grouped, temporal, and large datasets", while TabPFN-3.5 "substantially narrows these gaps". When the benchmark is restricted to tables of up to 100,000 rows, it says the model matches the best baselines on grouped and temporal data. It also notes that confidence intervals on the smaller slices are wide, so exact orderings there should be read cautiously. On speed, the core model runs up to twice as slow as TabPFN-3 on large training sets because its width was doubled; the Fast variant is up to three times faster than TabPFN-3, and up to six times faster than 3.5 itself, with a small loss of accuracy.

A grouped split means predicting for a group the model never saw in training, such as a new store or a new country. A temporal split means predicting later dates from earlier ones. If your problem is either of those, and your table is large, the report's own numbers say a well-tuned conventional model is still at least as good.

The routing below is ours, built on the limits and caveats the report states. Each row is a property of your data you can check before you run anything.

For the two most common business cases, churn and customer value, the input the model needs is the same table a conventional approach needs, which we set out in our guide to predictive analytics for CLV and churn. Demand forecasting is the case to treat with most care, because it is a temporal split by definition; our demand forecasting guide explains why a held-out later period is the only honest test. The report does note that the general 3.5 checkpoint, with no time-series fine-tuning, beat the previous time-series-specific TabPFN on a forecasting benchmark.

Section 4 of the report splits the release in two. The code, including the package that runs the models and the client SDK, is Apache 2.0. The weights for TabPFN-3.5 and TabPFN-3.5-Fast are under the TABPFN-3.5 License v1.0, which the report describes as permissive for academic use, research and evaluation in commercial settings. It explicitly allows testing, evaluation and internal benchmarking, so a company can download the model and try it on its own data without an agreement.

The restriction is equally explicit. The models, their derivatives and their outputs "cannot be used for commercial or production purposes", and the report lists what that includes: revenue-generating products, competitive benchmarking for procurement decisions, client deliverables, and using model outputs as inputs to internal commercial decision-making. That last item is broad. A dashboard that informs a pricing decision is, on the report's wording, a commercial use. For production the company offers a Commercial Enterprise License covering its managed API, SAP's generative AI hub, private-cloud deployment on AWS SageMaker and Azure AI Foundry, and custom deployments. The Plus and Thinking variants are only available under that licence.

For the agent builder, the appeal is mechanical. An analytics agent that has to answer "which of these accounts will churn" can call one endpoint with the table and get back a calibrated probability for each row, in one pass, with no training job to schedule and no model artefact to store. The report's description of cached inference, known rows processed once and new rows scored in under a second, is the property that makes it usable inside a conversational loop. Documentation for the package and the API is at Prior Labs' documentation. If you want that evaluation run on your own tables before you decide on a licence, our analytics team does it as a fixed-scope piece of work, with the grouped and temporal comparison the report itself recommends.