Release Notes

Version Mapping

Product Version Information

Product

MindSpeed

Product Version

26.0.0

Version Type

Official release

Component Name

MindSpeed LLM

Release Date

April 2026

Maintenance

6 months

Note

For version maintenance of MindSpeed LLM, see Version Maintenance Policy.

Table 1 MindSpeed LLM software version compatibility matrix

MindSpeed LLM version MindSpeed Core code branch name Megatron version PyTorch version TorchNPU version CANN version Python version
master (under development) master (under development) core_v0.12.1 2.7.1 In development In development Python 3.10
26.0.0 (commercial) 26.0.0_core_r0.12.1 core_v0.12.1 2.7.1 26.0.0 9.0.0 Python 3.10
2.3.0 (commercial) 2.3.0_core_r0.12.1 core_v0.12.1 2.7.1 7.3.0 8.5.0 Python 3.10
2.2.0 (commercial) 2.2.0_core_r0.12.1 core_v0.12.1 2.7.1 7.2.0 8.3.RC1 Python 3.10

Note

You can choose the MindSpeed LLM code branch as needed to download the source code and install it.

Version Compatibility Information

MindSpeed LLM version CANN version TorchNPU version
26.0.0 CANN 9.0.0
CANN 8.5.0
CANN 8.3.RC1
CANN 8.2.RC1
CANN 8.1.RC1
26.0.0
2.3.0 CANN 8.5.0
CANN 8.3.RC1
CANN 8.2.RC1
CANN 8.1.RC1
CANN 8.0.0
7.3.0
2.2.0 CANN 8.3.RC1
CANN 8.2.RC1
CANN 8.1.RC1
CANN 8.0.0
CANN 8.0.RC3
CANN 8.0.RC2
7.2.0

Version Usage Notes

None.

Update Notes

New Features

Component Description Purpose
MindSpeed LLM Added FSDP2 training support Supports Qwen3-30B, Qwen3-32B, Qwen3-235B, and Qwen3-Next model training
MindSpeed LLM Added 128K training support Supports ultra-long sequence training for gpt-oss and DeepSeekV3.2 models
MindSpeed LLM Improved tool efficiency Supports combined weight conversion and training, and combined data preprocessing and training
MindSpeed LLM Security hardening Supports PMCC protection for LLM fine-tuning

Removed Features

Component Description Purpose
MindSpeed LLM Model retirement Dense model retirement list:
Llama-2-34B
Llama-3-8B/70B
Llama-3.1-8B/50B/70B/200B
Llama-3.2-1B/3B
Llama-3.3-70B-Instruct
ChatGLM3-6B
GLM4-9B
Baichuan2-7B/13B
InternLM2.5-1.8B/7B/20B
Qwen2.5-0.5B/1.5B/3B/7B/14B/32B
Qwen3-8B (Megatron FSDP2)

MoE model retirement list:
Qwen3-30B (Megatron FSDP2)
GPT4-MoE-175B
Hunyuan-389B

API Changes

None.

Resolved Issues

None.

Known Issues

None.

Upgrade Impact

Impact on the Current System during Upgrade

  • Service impact.

    Upgrading the software version interrupts service.

  • Network communication impact.

    It has no impact on communication.

Impact on the Current System after Upgrade

None.

Document Summary Update Notes
MindSpeed LLM Installation Guide This guide helps you install MindSpeed LLM on an NPU. It covers hardware and operating system compatibility, driver firmware and CANN base software installation, and the complete installation process based on the PyTorch framework. It helps you quickly build a distributed LLM training environment. -
Quick Start: Qwen3-8B Model Pretraining and Fine-Tuning Using Qwen3-8B as an example, this guide helps developers who are new to MindSpeed LLM complete pretraining and fine-tuning tasks on the NPU. It helps you quickly get started with distributed LLM training. -

Virus Scan and Vulnerability Fix List

Virus Scan Results

Antivirus Software Name Antivirus Software Version Virus Database Version Scan Time Scan Result
QiAnXin 8.0.5.5260 2026-04-01 08:00:00.0 2026-04-02 No viruses, no malware
Kaspersky 12.0.0.6672 2026-04-02 10:05:00.0 2026-04-02 No viruses, no malware
Bitdefender 7.5.1.200224 7.100588 2026-04-02 No viruses, no malware

Vulnerability Fix List

None.