MindSpore
Huawei'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.
Overview
The framework carries an Apache 2.0 licence, which is good news and also the least interesting fact about it. What decides whether MindSpore belongs in a stack is whether you are buying Ascend. Huawei built it as the first-party framework for its own NPUs, and the graph compiler, operator fusion and memory planning all assume that target, which is why Ascend features land here months before they reach anywhere else.
Release activity is brisk once you look in the right place. Version 2.10.0 reached PyPI on 31 July 2026, following 2.7.2 in January, 2.8.0 on 30 January and 2.9.0 on 8 May. The GitHub mirror at mindspore-ai/mindspore still shows v2.3.0 from July 2024, so anyone judging the project from GitHub concludes it is abandoned. It is not. Development, issue triage and code review happen on Gitee, which means Mandarin in the pull request threads even though the English documentation on mindspore.cn is thorough.
On the CUDA question, be clear-eyed. Separate GPU wheels ended at mindspore-gpu 1.10.0 in March 2023, and the release notes for the newer mint API state that graph sinking mode and the CPU and GPU backends are not supported. Below the framework sits CANN, Huawei's Ascend compute stack, opened during 2025 under the CANN Open Software License Agreement version 1.0: a royalty-free, non-transferable and non-sublicensable grant limited to developing software for use with Ascend processors. That field-of-use restriction is not OSI open source and it needs legal review before it enters a shipping product. At the other end of the scale, MindSpore Lite is already installed on every HarmonyOS handset as a system AI engine covering CPU, NNRt and Kirin NPU, and Lite Micro extends to watch and earbud class silicon, which makes this the most credible path for anyone building consumer hardware inside Huawei's ecosystem.
Key Features
- ✓ Apache 2.0 on the framework itself, with no user cap or field-of-use clause, unlike the CANN stack beneath it
- ✓ Version 2.10.0 published to PyPI on 31 July 2026, following 2.7.2 in January, 2.8.0 on 30 January and 2.9.0 on 8 May the same year
- ✓ Ascend-native graph compilation with kernel fusion and memory planning aimed at Ascend 910 training clusters
- ✓ MindSpore Lite for on-device inference, shipped as the built-in AI engine in HarmonyOS across CPU, NNRt and Kirin NPU
- ✓ MindSpore Lite Micro targets microcontroller-class hardware such as smartwatches and earbuds, a class most frameworks ignore
- ✓ Imports MindIR, TensorFlow Lite, Caffe and ONNX models, so existing artefacts do not need retraining to be served
Where it holds
- • On Ascend it is the first-party path and gets kernel and feature support before any third-party framework does
- • Apache 2.0 on the framework, which is cleaner than the licences attached to most Huawei software
- • English documentation on mindspore.cn is unusually complete for a Chinese framework, covering API reference and tutorials
- • MindSpore Lite Micro reaches down to watch and earbud silicon, and MindSpore Lite is already resident on every HarmonyOS device
Where it breaks
- • The GitHub mirror stops at v2.3.0 from July 2024 while PyPI ships 2.10.0. Real development, issues and reviews happen on Gitee, in Mandarin
- • Separate CUDA wheels ended at mindspore-gpu 1.10.0 in March 2023, and the newer mint API is documented as not supporting the CPU and GPU backends. Treat this as Ascend-only in practice
- • CANN, the Ascend compute stack underneath, is released under the CANN Open Software License Agreement v1.0: a royalty-free but non-transferable grant limited to software for use with Ascend processors. That is a field-of-use restriction, not an OSI licence
- • Ecosystem gravity sits elsewhere. Porting a PyTorch training script is a rewrite and third-party model support lands late
- • Ascend hardware is export-controlled outside China, so a stack built on it is difficult to replicate for overseas customers
My Take
Judge this one by your procurement, not by its feature list. On Ascend it is the first-party stack and gets support before anything else, while the CUDA story has quietly lapsed: separate GPU wheels stopped at mindspore-gpu 1.10.0 in March 2023 and the newer mint API states plainly that CPU and GPU backends are not supported. Note also where the code actually lives. PyPI shipped 2.10.0 on 31 July 2026 while the GitHub mirror still shows v2.3.0 from July 2024, so the working repository is Gitee and the working language is Mandarin.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Framework is free under Apache 2.0. The cost is hardware: Ascend 910 training cards and Atlas inference modules, sold through Huawei channels and subject to export controls outside China. Huawei Cloud rents Ascend capacity, which needs a mainland or international Huawei Cloud account depending on region.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where MindSpore holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to MindSpore
PaddlePaddle
open-sourceBaidu'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.
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.