Paddle Releases
50 releases of PaddlePaddle/Paddle
- v2.37.0
The release introduces a new email builder, raises bulk send limits, and fixes several bugs related to mobile interactions and document editing.
Mar 4, 2026
- v3.3.0PaddlePaddle 3.3.0 Release Notes
PaddlePaddle 3.3.0 introduces major breakthroughs in large model training efficiency, development experience, and domestic hardware adaptation. Key improvements include FlashMaskV3 sparse attention optimization (2.1x faster than FlexAttention), FlexCheckpoint parameter auto-splitting system with AOA language, and virtual memory dynamic defragmentation reducing memory fragmentation from 10% to 3%.
Jan 31, 2026
- v0.1.1V0.1.1
MiroFish is a swarm intelligence engine that constructs high-fidelity parallel digital worlds to simulate future scenarios. This release highlights the project's workflow, integration with the OASIS engine, and provides comprehensive deployment instructions for both source code and Docker.
Jan 22, 2026
- v0.1.0V0.1.0
This is the initial release of MiroFish, introducing a universal swarm intelligence engine designed to predict anything by building parallel digital worlds. It establishes the core workflow for extracting reality seeds and simulating social evolution using autonomous agents.
Dec 22, 2025
- v3.2.2PaddlePaddle 3.2.2 Release Notes
PaddlePaddle 3.2.2 focuses on distributed training optimizations, operator mechanism improvements, and hardware adaptation enhancements to further improve framework performance and stability.
Dec 8, 2025
- RELEASE.2025-10-15T17-29-55ZSecurity/CVE RELEASE.2025-10-15T17-29-55Z
This is a security release addressing a Privilege Escalation vulnerability via Session Policy Bypass in Service Accounts and STS. It also includes various bug fixes regarding replication, LDAP TLS, pool decommissioning, and S3 POST policies.
Oct 16, 2025
- v3.2.0PaddlePaddle 3.2.0 Release Note
PaddlePaddle 3.2 brings significant improvements in large model training/inference performance, hardware adaptation, and ecosystem compatibility. Features include FlashMask V3 for storage-compute overlap, FP8 mixed precision training with 10-20% speed improvement, and comprehensive Safetensors format support.
Sep 8, 2025
- RELEASE.2025-09-07T16-13-09Z
A comprehensive release focusing on console updates, improving the idempotency of AbortMultipartUpload, adding Grafana dashboard variables, and fixing various issues related to Helm, LDAP, and OIDC authentication.
Sep 7, 2025
- v3.1.1PaddlePaddle 3.1.1 Release Note
PaddlePaddle 3.1.1 systematically strengthens the framework for large model training with operator precision fixes, FP8 optimization, and inference deployment improvements. Focuses on correctness and stability enhancements across the training pipeline.
Aug 21, 2025
- RELEASE.2025-07-23T15-54-02Z
This release introduces support for the X25519MLKEM768 post-quantum cryptographic algorithm. It also includes fixes for SinglePool restrictions and metadata read behavior.
Jul 23, 2025
- RELEASE.2025-07-18T21-56-31ZBugfix Release
A bugfix release addressing issues with Console updates, CopyObject checksum preservation, Admin API stability, and Lambda handler response matching.
Jul 18, 2025
- v3.1.0PaddlePaddle 3.1.0 Release Note
PaddlePaddle 3.1 focuses on automatic parallel architecture refinement, FP8 low-precision training support (10-20% speed improvement), and heterogeneous multi-chip adaptation. Provides CUDA-like operator reuse mechanism reducing hardware adaptation costs.
Jun 29, 2025
- RELEASE.2025-06-13T11-33-47ZBugfix Release
This is a bugfix release addressing several S3 compatibility issues, particularly regarding ILM (Lifecycle Management) and multipart uploads. It also introduces RISC-V 64 cross-compilation support and updates bucket replication metrics.
Jun 23, 2025
- v3.0.0PaddlePaddle 3.0.0 Release Note
PaddlePaddle 3.0 is a major release introducing unified dynamic-static automatic parallelism, neural network compiler CINN with 27.4% average performance improvement, high-order automatic differentiation for scientific computing 115% faster than PyTorch, and training-inference unification with 95% dynamic-to-static conversion success rate.
Mar 31, 2025
- v3.0.0-beta0PaddlePaddle 3.0.0-beta0 Release Note
PaddlePaddle 3.0 Beta introduces unified dynamic-static automatic parallelism, neural network compiler CINN with 30% performance improvement on Llama2 and Stable Diffusion, and high-order automatic differentiation 70% faster than competitors.
Jun 27, 2024
- v2.6.1PaddlePaddle 2.6.1 Release Note
PaddlePaddle 2.6.1 is a maintenance release adding Fake GroupWise Quantization support, PGLBox graph neural network training engine, and multiple bug fixes including security patches.
Apr 9, 2024
- v2.6.0PaddlePaddle 2.6.0 Release Note
PaddlePaddle 2.6.0 introduces PIR (Paddle Intermediate Representation), upgraded dynamic-to-static with 100% success rate, enhanced distributed training, and comprehensive API improvements including 109 new inplace APIs.
Jan 8, 2024
- v2.5.0PaddlePaddle 2.5.0 Release Note
PaddlePaddle 2.5.0 introduces unified dynamic-static architecture with CINN compiler, PHI operator library unification, new static graph executor, and 0-dim tensor support.
Jul 25, 2023
- v2.4.2PaddlePaddle 2.4.2 Release Note
This release focused on fixing known bugs and adding minor features. Key improvements included DLpack object handling, multiplex API fixes, and sparse convolution performance via Cutlass. In the deployment direction, mixed-precision inference was enhanced, and Jetson Ampere architecture support was added.
May 8, 2023
- v2.4.1PaddlePaddle 2.4.1 Release Note
This version removed the dependency on `python.so` to resolve execution failures in specific environments, including conda.
Dec 13, 2022