3.1.3 Patch Release (Jan 08 2026)

  • Scikit-learn 1.8 compatibility fix (#11858)

  • Add ARM CUDA wheels for PyPI. (#11827) Add nccl as dep for aarch64. (#11753)

  • [R] Fix off-by-one bug: nrounds=0 resulted in 2 iterations #11856

  • [R] Fix mingw warnings, winbuilder check warnings, memory safety issues. (#11859, #11847, #11830, #11906)

  • Avoid overflow in rounding estimation. (#11910)

  • Workaround compiler issue on Windows, affects the use of max_delta_step with CUDA. (#11916)

3.1.2 Patch Release (Nov 20 2025)

  • Fix loading nccl 2.28. (#11806)

  • Fix ordering of Python callbacks. (#11812)

  • Infer the enable_categorical during model load. (#11816)

3.1.1 Patch Release (Oct 22 2025)

  • Emit correct error when performing inplace-predict using a CPU-only version of XGBoost, but with a GPU input. (#11761)

  • Enhance the error message for loading the removed binary model format. (#11760)

  • Use the correct group ID for SHAP when the intercept is a vector. (#11764)

3.1.0 (2025 Sep 22)

We are delighted to share the latest 3.1.0 update for XGBoost.

Categorical Re-coder

This release features a major update to categorical data support by introducing a re-coder. This re-coder saves categories in the trained model and re-codes the data during inference, to keep the categorical encoding consistent. Aside from primitive types like integers, it also supports string-based categories. The implementation works with all supported Python DataFrame implementations. (#11609, #11665, #11605, #11628, #11598, #11591, #11568, #11561, #11650, #11621, #11611, #11313, #11311, #11310, #11315, #11303, #11612, #11098, #11347) See Auto-recoding (Data Consistency) for more information. (#11297)

In addition, categorical support for Polars data frames is now available (#11565).

Lastly, we removed the experimental tag for categorical feature support in this release. (#11690)

External Memory

We continue the work on external memory support on 3.1. In this release, XGBoost features an adaptive cache for CUDA external memory. The improved cache can split the data between CPU memory and GPU memory according to the underlying hardware and data size. (#11556, #11465, #11664, #11594, #11469, #11547, #11339, #11477, #11453, #11446, #11458, #11426, #11566, #11497)

Also, there’s an optional support (opt-in) for using nvcomp and the GB200 decompression engine to handle sparse data (requires nvcomp as a plugin) (#11451, #11464, #11460, #11512, #11520). We improved the memory usage of quantile sketching with external memory (#11641) and optimized the predictor for training (#11548). To help ensure the training performance, the latest XGBoost features detection for NUMA (Non-Uniform Memory Access) node (#11538, #11576) for checking cross-socket data access. We are working on additional tooling to enhance NUMA node performance. Aside from features, we have also added various documentation improvements. (#11412, #11631)

Lastly, external memory support with text file input has been removed (#11562). Moving forward, we will focus on iterator inputs.

Multi-Target/Class Intercept

Starting with 3.1, the base-score (intercept) is estimated and stored as a vector when the model has multiple outputs, be it multi-target regression or multi-class classification. This change enhances the initial estimation for multi-output models and will be the starting point for future work on vector-leaf. (#11277, #11651, #11625, #11649, #11630, #11647, #11656, #11663)

Features

  • Support leaf prediction with QDM on CPU. (#11620)

  • Improve seed with mean sampling for the first iteration. (#11639)

  • Optionally include git hash in CMake build. (#11587)

Removing Deprecated Features

This version removes some deprecated features, notably, the binary IO format, along with features deprecated in 2.0.

  • Binary serialization format has been removed in 3.1. The format has been formally deprecated in 1.6. (#11307, #11553, #11552, #11602)

  • Removed old GPU-related parameters including use_gpu (pyspark), gpu_id, gpu_hist, and gpu_coord_descent. These parameters have been deprecated in 2.0. Use the device parameter instead. (#11395, #11554, #11549, #11543, #11539, #11402)

  • Remove deprecated C functions: XGDMatrixCreateFromCSREx, XGDMatrixCreateFromCSCEx. (#11514, #11513)

  • XGBoost starts emit warning for text inputs. (#11590)

Optimizations

  • Optimize CPU inference with Array-Based Tree Traversal (#11519)

  • Specialize for GPU dense histogram. (#11443)

  • [sycl] Improve L1 cache locality for histogram building. (#11555)

  • [sycl] Reduce predictor memory consumption and improve L2 locality (#11603)

Fixes

  • Fix static linking C++ libraries on macOS (#11522)

  • Rename param.hh/cc to hist_param.hh/cc to fix xcode build (#11378)

  • [sycl] Fix build with updated compiler (#11618)

  • [sycl] Various fixes for fp32-only devices. (#11527, #11524)

  • Fix compilation on android older than API 26 (#11366)

  • Fix loading Gamma model from 1.3. (#11377)

Python Package

  • Support mixing Python metrics and built-in metrics for the skl interface. (#11536)

  • CUDA 13 Support for PyPI with the new xgboost-cu13 package. (#11677, #11662)

  • Remove wheels for manylinux2014. (#11673)

  • Initial support for building variant wheels (#11531, #11645, #11294)

  • Minimum PySpark version is now set to 3.4 (#11364). In addition, the PySpark interface now checks the validation indicator column type and has a fix for None column input. (#11535, #11523)

  • [dask] Small cleanup for the predict function. (#11423)

R Package

Now that most of the deprecated features have been removed in this release, we will try to bring the latest R package back to CRAN.

  • Implement Booster reset. (#11357)

  • Improvements for documentation, including having code examples in XGBoost’s sphinx documentation side, and notes for R-universe release. (#11369, #11410, #11685, #11316)

JVM Packages

Documents

Changes for general documentation:

CI and Maintenance