Last chance to register for the TabPFN-3.5 Model Deep Dive session coming up on the 28th! Join @FrankRHutter and @tuanacelik for an overview of what's changed with our newest model release, what the Fast, Plus and Thinking modes bring to the table, benchmark results, as well as a mini hands on demo! We'll have some time for questions at the end too, as well as some updates about what to expect from us in the coming weeks! 📅 Monday, September 28, 5:00 pm CET 🔗 Register: app.livestorm.co/p/77a95949-…

A state-of-the-art model that has never seen a real dataset. Our co-CEO @FrankRHutter sat down with @MLStreetTalk on why deep learning kept failing on tables, and how TabPFN learns the algorithm itself from hundred of millions of synthetic datasets. Full episode: piped.video/watch?v=72Im-Mm5…

Most real tables are a mix of categoricals, free text, missing values..the list can go on. So to see how TabPFN-3.5 fares, we created a cookbook predicting a restaurant's Zomato rating (0 to 5) from its listing using a dataset of 41,665 restaurants. What the model sees per restaurant: · Categoricals: location, cuisine types, restaurant type (Quick Bites, Casual Dining...), whether it takes online orders or table bookings · Numeric: number of votes, approximate cost for two · Free text: the full list of customer reviews, the menu items, and the dishes people liked A few things to take from this: 🔥 Thinking mode helps both versions. Set thinking_effort="high" and a thinking_metric (RMSE here) and the model spends extra compute at fit time to get the best results possible. 3.5-Thinking is the best model in the table. 💬 Text is the differentiator. Reviews carry most of the signal about a rating, and TabPFN reads them natively. XGBoost gets TF-IDF features and lands at ~3x the error. Try out the cookbook here: docs.priorlabs.ai/cookbook/z…

Just a few days to go. We're hosting an unconference with @deepset_ai, makers of Haystack in Berlin on Wednesday, September 24th. Open agents meet open tabular models. Unconference format means the agenda is shaped by what you bring. If you're working with structured data, building agentic systems in production, or figuring out how to own your AI stack end to end, this is for you. 📍 deepset HQ, Berlin 🗓️ September 24th, 18:00–20:30 Last spots available: luma.com/haystack-priorlabs

TabPFN-3.5 is out, and it now ranks first on both TabArena and BeyondArena. On September 28 we’re hosting a model deep dive with @FrankRHutter, to walk through what actually changed and why it matters. A few things we'll get into: · Where 3.5 improves most: non-IID data. Grouped and temporal rows, high-cardinality and high-dimensional features, and tables where the columns are messy or mixed. · The full model family and when to reach for each one: Base for local inference, Plus for enhanced text processing on the API, Thinking for maximum accuracy and the new Fast alpha with up to 6x faster inference. · The numbers behind the leaderboard placements. · What's coming next from Prior Labs and the community. 📅 Monday, September 28, 5:00 pm CET 🔗 Register: app.livestorm.co/p/77a95949-…

We ran TabPFN-3.5 against TabPFN-3, TabFM, and tuned XGBoost on OpenML's Taiwanese Bankruptcy dataset. 6,819 rows, 94 features. TabPFN-3.5 scored 0.9571 ROC AUC. TabPFN-3 scored 0.9506. TabFM 0.9507 and tuned XGBoost scored 0.9499. ROC AUC measures how well each model ranks bankrupt companies above non-bankrupt ones (higher = better). We compare it with log loss to assess the quality of the predicted probabilities, and with latency to judge the cost of making predictions. In this experiment, TabPFN-3.5 had the highest ROC AUC on the split we used, and lowest log loss, however XGBoost has much lower fit-plus-predict latency. 🔗 You can run the the whole thing yourself with this cookbook: docs.priorlabs.ai/cookbook/t…

TabPFN-3.5-Plus is now available in SAP AI Core. SAP customers can run it directly and start predicting on their business data without leaving their environment. No GPUs to spin up, no separate infrastructure, your data stays where it is. Create a deployment, point it at your table, predictions come back. And it follows the same process as other models in SAP AI Core. Get the full walkthrough in the blog: community.sap.com/t5/technol…

TabPFN-3.5 is here and it's the fastest and most accurate tabular foundation model, ranking #1 on TabArena and BeyondArena! So to celebrate, we're hosting a hackathon 🎉 Build anything with TabPFN-3.5 and win an Nvidia GPU. It runs through October 6th. Three prizes: 🥇 DGX Spark 🥈 Jetson AGX Orin 🥉 GeForce RTX 4090 Show us what you can do. Agents, mode comparisons, extensions, whatever you want. Literally.. Submit a repo, add a demo video for a chance to win! Join here: platform.priorlabs.ai/hackat…

We’re hosting a hackathon to celebrate the release of TabPFN-3.5! Join for a chance to win one of 3 Nvidia GPUs: 🥇 DGX Spark 🥈 Jetson AGX Orin 🥉 GeForce RTX 4090 The challenge is simple: showcase TabPFN-3.5 in any shape of form you desire. • Build agents using TabPFN-3.5 for predictions • Showcase the different modes (thinking, plus or fast) • Build extensions • The world is your oyster 🦪 You have until the 6th of October! Requirements: Submit a runnable repository with an optional video showcasing your submission and demo! Good luck! Join via the Prior Labs platform: platform.priorlabs.ai/hackat…

TabPFN-3.5 is here. It is the most accurate and scalable tabular foundation model available today: first place on TabArena and BeyondArena, ahead of any other model while being up to 20x faster than the next places. Real data is rarely clean and independent: insurance claim data combines the amount with reviewer conclusions, transactions tied to merchants and customers, readings of hundreds of correlated sensors. TabPFN-3.5 improves most on exactly this data: text-rich, grouped, high-cardinality and high-dimensional. And we really mean it with high-dimensional: up to 20k columns supported (6k recommended), a 10x jump from a 2,000 limit we had with TabPFN-3. Also today: an updated Thinking mode for when accuracy matters more than compute, and TabPFN-3.5-Fast in alpha, up to 6× faster than the base model. Available now via our API, SAP AI Core, AWS SageMaker and the open-source tabpfn package. To let you test it at its fullest, you get 2x token calls on our API for the next 2 weeks. Get started now: docs.priorlabs.ai/changelog/…

We're hosting an unconference in Berlin with @deepset_ai on open agents and open tabular models. Agents have evolved past single LLMs doing everything. They manage reasoning loops, tool selection, state... TabPFN handles structured data prediction tasks where generative models fall short. We'll spend a whole evening discussing these topics. Come join us for an evening of discussions and networking Topics we expect to cover: → Structured data in agentic workflows → When agents should hand off to TabPFN for prediction → Benchmarking tabular models vs LLMs → Why orchestration matters for production agents → What owning your stack end to end actually requires → Open-weight vs closed models in practice 📍 deepset HQ, Berlin 🗓️ September 24th, 18:00–20:30 Registration link: luma.com/haystack-priorlabs

Forecasts you cannot explain rarely help you make decisions. Our new cookbook forecasts German day-ahead electricity prices and then shows three ways to understand what drove the result using TabPFN time-series forecasts with partial dependence, Window SHAP, and additive decomposition. 1️⃣ Partial dependence plots show how the average forecast shifts as one input moves. 2️⃣ Window SHAP gives you feature-by-time attribution. Each cell in the heatmap shows which features pushed that forecast window up or down. 3️⃣ Additive decomposition splits the observed series into trend, hour-of-day seasonality, day-of-week seasonality, and residuals. The example uses seven days of context and predicts the final 24 hours. All three methods are included in the tabpfn-time-series explainability repository. Get started: docs.priorlabs.ai/cookbook/t…