TabPFN-3.5

TabPFN-3.5 improves most on real-world prediction tasks, where the incoming ERP, CRM and production data isn’t clean, with the model family ranking first on TabArena and BeyondArena. Plus and Thinking are updated to version 3.5 as well and deliver additional gains on grouped, temporal and text-rich data. We also introduce an alpha fast inference model, TabPFN-3.5-Fast.

​

Model

TabPFN-3.5 is our first multitask model, improving both classification and regression within a single checkpoint. It achieves state-of-the-art performance on BeyondArena’s wide datasets and improves on high-cardinality categories, reducing the need for manual encoding and feature selection. With that, we also increase the maximum number of features supported tenfold to 20,000, with up to 6,000 recommended for better performance. TabPFN-3.5 also improves predictive distributions for skewed and zero-heavy regression targets such as claim amounts and customer spend. It ranks first on ScoringBench, which evaluates the quality of predictive distributions. Building on the base model, TabPFN-3.5-Plus significantly improves performance on text-rich data, making it well-suited for tasks where signal comes from product descriptions, reviews or inspection notes. TabPFN-3.5-Thinking, our most accurate model, further improves on grouped and temporal data. This helps with tasks where rows belong to related entities or observations follow a time sequence. Applications include fraud scoring across accounts, churn prediction from customer histories, and demand forecasting across stores and products. TabPFN-3.5-Fast, available in alpha, delivers up to 6× faster inference than the base model for latency-sensitive applications. See the TabPFN-3.5 model family page and our technical report for more details.

---

​

Open-source package

TabPFN-3.5 and TabPFN-3.5-Fast (alpha) model weights are available for local inference and research starting today. The package downloads the weights after a one-time license acceptance. Check the access guide for setup instructions and model licensing for usage terms. The 9.0.0 release of the tabpfn open-source package keeps the existing interfaces. This release also adds:

• Built-in text and datetime preprocessing. Local inference can process text columns and datetime columns directly into features, reducing the time spent on feature engineering.
• Inference defaults. To improve performance, we have introduced model-specific inference configurations. With that, softmax_temperature now defaults to "auto". Both it and n_estimators read their defaults from the selected checkpoint.
• Keep rare outcomes when subsampling. The new majority_downsample option keeps all non-majority rows while sampling the most frequent target value, including zeros in regression.
• Smaller KV caches for TabPFNClassifier. Classification caches now store projected decoder keys instead of full embeddings. See KV cache.
• More efficient processing of large numeric tables. We optimized RAM and GPU memory across preprocessing, fit() and predict(), with 41% lower peak RAM in our large-table benchmark.

We also shipped:

• Categorical feature declarations in string columns are now respected regardless of how many distinct values a column contains. The declaration can now also be made through the pandas category dtype.
• Saved models now load with device="auto" by default (#1243).
• Fine-tuning now chunks large batches to prevent a memory-access crash (#1184).
• Classes absent from the validation set now receive neutral thresholds during metric tuning (#1152).

See Migrating from TabPFN-3 for behavior changes.

---

​

tabpfn-client and API

TabPFN-3.5-Plus improves results on text-rich tables by adding proprietary text processing to the base model. It ranks #1 on STRABLE, a benchmark for tabular predictions over tables with strings, and is available through the hosted API, tabpfn-client, MCP, and integrations. TabPFN-3.5-Thinking is our most accurate offering, especially on grouped and temporal data. It uses more compute during fitting to improve prediction quality, and the fitted model can then be reused for repeated predictions. Thinking is available through the hosted API, tabpfn-client, and enterprise deployments; see Models for its limits. TabPFN-3.5-Fast (alpha) and Base models are also available through the API and can be selected through the TabPFN instance constructor. We have also updated API metering, which now accounts for the selected model and operation performed, including 75% lower charges for successful KV cache reuse. Thinking fits and predictions share the same token budgets, replacing the separate Thinking fit limit. From 15 September until 29 September 2026, all API and MCP users get a 50% reduction from the standard TabPFN-3.5 token rates for the 3.5 family. Standard rates resume on 29 September. The 0.6.0 release of tabpfn-client adds:

• Estimate token cost before running a request. estimate_cost() sends only dataset dimensions and settings and returns cost in tokens, without uploading data or consuming usage limits. See API metering for more details.
• Reuse fitted models across Python sessions. save_model() and load_model() let you resume predictions without fitting again. The saved record references the server-side model; use the same account. See Thinking mode and KV cache for more details.
• Inspect prediction uncertainty visually. plot_regression_distribution() plots the predicted regression distribution. Install tabpfn-client[viz] to use it.

We also shipped:

• fit() and predict() no longer print progress spinners, and routine messages use debug logging (#385).
• Restored compatibility with scikit-learn tools such as partial dependence (#372).

---

​

Integrations

SAP AI Core is a new way to access TabPFN-3.5-Plus, available today to all SAP customers. TabPFN-3.5-Plus and Fast (alpha) are available on AWS SageMaker, with Thinking available through a private listing. TabPFN-3.5-Plus is also available on Microsoft Foundry. The TabPFN-3.5 model family is also available as an on-premises container. For enterprise upgrades, container access or private listings, contact our sales team.

---

​

Migrating from TabPFN-3

For both the local package and API:

1. Compare TabPFN-3.5 with TabPFN-3 on your validation set. Test raw text and categorical columns before removing existing encoders or feature selection. See Preprocessing.
2. Declare categorical identifiers explicitly so they are not mistaken for numerical or free-text features. See Feature engineering.
3. Fit a new model when switching versions. Existing fitted models and KV caches remain tied to the version used to create them.
4. Check the model limits for your selected variant, as well as for text feature limits.

​

Open-source package

1. Upgrade with pip install --upgrade tabpfn and complete the weight-access setup.
2. Select ModelVersion.V3_5 for the base model or ModelVersion.V3_5_FAST for Fast through create_default_for_version(). See Selecting Model Version.
3. For built-in text processing, use pandas string or pyarrow string columns. Undeclared string columns with more than 30 distinct values qualify as text.
4. Keep dates in a datetime dtype for them to be processed correctly.
5. Keep the selected model’s defaults when customizing inference. Pass only your changes in an inference_config dictionary. Passing a full InferenceConfig object is still supported but is deprecated.
6. If you extract training embeddings, refit with fit_mode="fit_preprocessors". Training embeddings are no longer available with fit_with_cache; test embeddings remain available.
7. Existing fine-tuned checkpoints remain based on TabPFN-3. To use the new model, repeat fine-tuning with TabPFN-3.5. Fine-tuning estimators still require text and datetime features to be converted beforehand.

​

API client

1. Upgrade with pip install --upgrade tabpfn-client.
2. Select Plus with create_default_for_version("v3.5"), or Fast with create_default_for_version("v3.5-fast"). Use "v3" to keep using TabPFN-3-Plus. See Selecting Model Version.
3. For TabPFN-3.5-Thinking, thinking_mode=True selects the 3.5 version. If you’d like to use TabPFN-3-Thinking, pass "v3" as the model version.
4. Review the updated API metering. Token charges depend on the selected model and settings.

​

REST API

1. Select model_path="v3.5_default" or "v3.5-fast_default" inside tabpfn_config - see Selecting Model Version for more details.
2. The upload, fit, and predict flow is unchanged.
3. Check GET /tabpfn/get_model_limits for the API’s current limits.

​

Enterprise deployments

Deploy the updated image or marketplace model to use TabPFN-3.5. Existing endpoints do not upgrade automatically. Follow the SageMaker or Foundry deployment guide for your platform.

---

​

Get started