TabPFN-3.5
#2 in LLM-Inferenz & ServingPrior Labs · v3.5 · 3× · tolest 19. Sept. 2026
TabPFN-3.5 is a tabular Foundation Model by Prior Labs (part of SAP since July 2026) that solves classification and regression for structured data in a single forward pass. Version 3.5 is the first multitask checkpoint of the model family, supports up to 20,000 features, improves processing of text, high-cardinality categories, and wide tables, and introduces an additionally latency-optimized Alpha variant "Fast" as well as updated "Plus" and "Thinking" variants. According to the provider, it ranks first on the Benchmarks TabArena and BeyondArena and is available via hosted API, Python client, MCP, AWS SageMaker, Microsoft Foundry, SAP AI Core, as well as an On-Premises container. The model weights are available under a non-commercial license.
Features
| Deployment (Self-Hosted/Cloud) | Self-hosting via on-premises container/open-source package, private cloud, plus managed cloud via AWS SageMaker, Microsoft Foundry and SAP AI Core |
| Throughput/Latency | Cached inference scores new rows in well under a second; supports up to 1,000,000 training rows; 75% lower charges for successful KV cache reuse |
| License | Non-commercial TabPFN-3.5 license for open model weights; API/Private Cloud governed by separate Terms of Service (Plus/Thinking line) |
| Platform | Hosted API, tabpfn-client (Python), MCP, AWS SageMaker, Microsoft Foundry/Azure ML, SAP AI Core, on-premises container |
| Price | Model weights free (non-commercial license); API/MCP: 50% discount on standard token rates Sep 15–29, 2026; commercial on-prem/private cloud use requires a license from Prior Labs |
| Protocol Compatibility | Available via hosted REST API, tabpfn-client Python package, and Model Context Protocol (MCP) |
| Release Date | September 15, 2026 (discount announcement start date); model checkpoint dated 2026-09-09 |
| Supported Models/Providers | Model family: Base, Fast (alpha), Plus, Thinking – single multitask checkpoint for classification & regression |