MindSDK
MindSDK delivers various AI software development kits (SDKs) accelerated by Ascend AI Processors and provides simplified APIs to accelerate the development of high-performance AI applications for thousands of industries.
Key Features of MindSDK 26.1.0
The key updates in this MindSDK version are as follows:
- Rec SDK: The training framework and generative recommendation operators are adapted to
Ascend 950 products . The torch_rec_v2 module introduces new features such as dynamic table eviction policies and on-chip memory and DDR cache. - Index SDK: Supports filtering of temporal and spatial attributes of TS FlatIP and Int8Cos in A2/A3 scenarios, supports the IVF-RabitQ algorithm, and introduces new APIs such as train and remove_ids.
- RAG SDK: Accelerates BGE embedding and reranker models, with a focus on optimizing the performance of bge-reranker-v2-m3 and bge-m3.
- Multimodal SDK: Adds examples for key frame filtering, automatic optimization, and multi-scale resampling, as well as the reference design for SCC visual token compression.
- Agent SDK: Aura provides an end-to-end Agentic reinforcement learning training framework, compatible with multiple models ranging from qwen3-4b to qwen3-32b.
- Driving SDK: All operators and mainstream autonomous driving/VLA models are fully supported by
Ascend 950 products , and new models such as GR00T-N1.6 are added.
For details about the complete version mapping, version compatibility, updates, and upgrade impact, see the Release Notes.