TabPFN-3-Plus opens enterprise API

Tabular ML is the boring, high-value part of AI — credit scores, churn models, fraud checks, clinical risk, pricing, forecasting. It runs huge parts of real businesses. But getting strong tabular models into production has usually meant a lot of tuning, glue code, and platform work. Prior Labs is trying to make that much simpler. On May 12, it released TabPFN-3 and opened TabPFN-3-Plus as an API and enterprise product for private deployments, including AWS SageMaker VPC setups and Azure AI Foundry. ### What is TabPFN, exactly? TabPFN is a foundation model for structured tables rather than text or images. Instead of training a fresh model from scratch for every spreadsheet-shaped problem, it uses a pretrained transformer to do classification and regression directly on tabular data. Prior Labs pitches that as “state-of-the-art predictions in seconds” with far less tuning than the old default stack of boosted trees and AutoML pipelines. (priorlabs.ai) ### What changed this week? The actual news is two-part. First, Prior Labs published TabPFN-3 as its new flagship model. Second, it packaged the stronger “Thinking” variant, TabPFN-3-Plus, as something enterprises can buy and deploy through an API, on-prem, or inside private cloud environments. That includes AWS SageMaker and Azure AI Foundry, which is the part enterprise buyers care about because it drops the model into tooling they already use. (docs.priorlabs.ai) ### Why is “3-Plus” different from plain TabPFN-3? TabPFN-3 is the base release. TabPFN-3-Plus adds a “Thinking mode,” which basically spends more compute during fitting to get better results. Prior Labs says that mode improves performance by up to 15% on its TabArena ELO metric over vanilla TabPFN-3. In the technical report, it says 3-Plus beats all non-TabPFN models by more than 200 Elo on TabArena, and by 420 Elo on the largest subset. (priorlabs.ai) ### Are the weights really open? Not in the usual “open weights for production” sense. This is the catch. Prior Labs says it is releasing TabPFN-3 weights and inference code under a TabPFN-3 license that is permissive for research and internal evaluation. The Hugging Face license text is more explicit — non-commercial and non-production use only. If you want production access through the API, separate commercial terms apply. (docs.priorlabs.ai) ### Why does SageMaker support matter? Because model quality is only half the sale. Enterprises want the model to live inside their own cloud account, touch governed data, and fit into existing model catalogs and endpoints. Prior Labs already had SageMaker materials in its repo, including an example for invoking a TabPFN endpoint on AWS. Now it is clearly positioning TabPFN-3-Plus as a first-class enterprise deployment option rather than a research demo with a notebook attached. (priorlabs.ai) ### How big is the performance claim? Pretty aggressive. Prior Labs says TabPFN-3 is up to 20x faster than TabPFN-2.5, can scale to 1M rows on a single H100 using row-chunking and a reduced KV cache, and that 3-Plus can outperform AutoGluon 1.5 extreme in less than one-tenth the runtime on its benchmark framing. That is the pitch in one sentence — strong tabular accuracy without the usual slow tuning loop. ### So what’s the real takeaway? (priorlabs.ai) This is less an “open model drop” than a go-to-market shift. Prior Labs is turning TabPFN from a beloved research tool into something procurement and platform teams can actually adopt. If the benchmarks hold up in customer workloads, that is a real wedge into enterprise tabular ML. But the company is still keeping the highest-value production path behind API and enterprise licensing. (priorlabs.ai)