PaddlePaddle
Baidu's training and inference framework (飞桨), Apache 2.0, with kernels for Kunlunxin XPU, Ascend NPU, Hygon DCU and Cambricon MLU alongside CUDA. PaddleOCR is why most teams arrive.
Overview
Baidu opened Paddle in 2016 and it has since become the default teaching framework across Chinese universities. That matters more than it sounds: hiring an engineer in Shenzhen who already writes Paddle is easy in a way that hiring for JAX is not. Framework and model suites are Apache 2.0 with no extra terms. Version 3.4.0 was tagged on GitHub on 12 June 2026 while PyPI wheels stand at 3.3.1 from 24 March 2026, so check which artefact you are actually installing before you pin.
The reason to care is the hardware abstraction. The 3.0 line rebuilt the adaptation layer so a single codebase compiles to NVIDIA CUDA, Kunlunxin XPU, Ascend NPU, Cambricon MLU or Hygon DCU, and PaddleX surfaces that as a switch at pipeline level rather than a fork of your code. If your buyer is a state-owned enterprise or any organisation working under a domestic silicon mandate, that capability decides whether you can bid at all. Quality across those targets is not equal, with Kunlunxin visibly getting the most attention, so validate on the exact card before you promise numbers.
Most teams arrive through the model suites rather than the framework itself. PaddleOCR remains the strongest open Chinese-language OCR available, and PaddleOCR-VL 1.6, released on 28 May 2026 at roughly 0.9B parameters, reports 96.33 on OmniDocBench v1.6 with improvements aimed at tables, seals and rare characters. Edge deployment is the weaker half of the story: Paddle Lite converts models to .nb NaiveBuffer files for ARM CPU and OpenCL, but its last tag was v2.14-rc in August 2024. For on-device work in 2026, plan on exporting and running the model in another runtime.
Key Features
- ✓ Apache 2.0 across the framework and the model suites, with no separate community licence layered on top
- ✓ v3.4.0 tagged on GitHub 12 June 2026, while PyPI wheels currently sit at 3.3.1 from 24 March 2026, a gap worth checking before pinning
- ✓ Hardware adaptation layer letting one codebase target NVIDIA CUDA, Kunlunxin XPU, Ascend NPU, Cambricon MLU or Hygon DCU, exposed at pipeline level through PaddleX
- ✓ CINN neural network compiler and unified automatic parallelism introduced in the 3.0 line for large model training
- ✓ PaddleOCR, PaddleDetection, PaddleSeg and PaddleNLP suites with pretrained weights; PaddleOCR-VL 1.6, released 28 May 2026 at roughly 0.9B parameters, reports 96.33 on OmniDocBench v1.6
- ✓ Paddle Lite compiles models to .nb NaiveBuffer files for ARM CPU and OpenCL GPU inference on phones and IoT boards
Where it holds
- • The domestic silicon story is real and no other framework matches it: Kunlunxin, Ascend, Cambricon and Hygon are all first-party targets, which decides tenders in China
- • PaddleOCR is the strongest open OCR stack for Chinese text anywhere, and the VL models are small enough to serve on one GPU
- • Taught in Chinese universities, so hiring engineers in Shenzhen or Hangzhou who already know the API is straightforward
- • Apache 2.0 throughout, weights included, so nothing blocks commercial redistribution
Where it breaks
- • Paddle Lite, the on-device runtime, last tagged v2.14-rc in August 2024. The edge path has drifted while the server framework ships every few months
- • API documentation is Chinese-first and English pages routinely trail a version behind, which bites hardest on newer distributed and compiler features
- • Porting a PyTorch codebase is a rewrite, not a shim. The API resembles torch without being compatible with it
- • Accelerator support is uneven in practice. Kunlunxin gets the most engineering attention, Cambricon and Hygon noticeably less
My Take
Written off outside China as a torch clone with worse English docs, which misses what it is for. The value sits in the hardware layer: one codebase compiling down to Kunlunxin XPU, Ascend NPU, Hygon DCU or Cambricon MLU, which is the difference between winning and losing a tender where domestic silicon is a procurement requirement. PaddleOCR alone justifies the install for anyone handling Chinese documents, with PaddleOCR-VL 1.6 at around 0.9B parameters reporting 96.33 on OmniDocBench v1.6 in May 2026. The soft spot is the edge, where Paddle Lite has not tagged a release since v2.14-rc in August 2024.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Framework and model suites are free under Apache 2.0. Baidu sells hosted training and serving through Baidu AI Cloud, which does require a mainland account and, for most enterprise quotas, a Chinese business licence. Nothing in the open source path requires either.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where PaddlePaddle holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to PaddlePaddle
MindSpore
open-sourceHuawei's Apache 2.0 framework (昇思) built around Ascend NPUs. Development is on Gitee, the GitHub mirror is stuck on 2024 and MindSpore Lite is the built-in AI engine in HarmonyOS.
Huawei Cloud ModelArts
paidHuawei Cloud's full-lifecycle AI platform. Training and serving on Ascend NPUs through CANN, with Lite Server, Lite Cluster and a MaaS layer. Proprietary and quote-priced.