Deploy DeepSeek‑V3 on Ascend: Step‑by‑Step Guide for Fast AI Inference

DeepSeek AI recently released the multimodal large model Janus‑Pro and earlier models DeepSeek‑R1, V3, and V2, which support the Ascend platform for efficient inference.

On 2025‑02‑04 the models were made available on the Ascend community, allowing one‑click download and immediate deployment on Ascend hardware.

Hardware Requirements

Running DeepSeek‑V3 requires four Atlas 800I A2 (8×64 GB) servers.

Weight Conversion

Convert the model weights for GPU and NPU as shown in the following diagrams:

Load the MindIE Image

Download the MindIE Docker image adapted for DeepSeek‑V3, e.g. mindie:1.0.T71-800I-A2-py311-ubuntu22.04-arm64, and verify it with docker images.

Start the Container

Place the converted weights in the model directory, set the folder ownership to 1001 and permissions to 750, then start the container.

Service‑Level Testing

Enable the expandable_segments environment variable to activate the memory‑pool extension feature, adjust service parameters as needed, and launch the service. Successful start is indicated by “Daemon start success!”.

For more details and documentation, visit the Ascend model zoo pages for DeepSeek‑R1, DeepSeek‑V3, and Janus‑Pro.

https://www.hiascend.com/software/modelzoo/models/detail/68457b8a51324310aad9a0f55c3e56e3

https://www.hiascend.com/software/modelzoo/models/detail/678bdeb4e1a64c9dae51d353d84ddd15

https://www.hiascend.com/software/modelzoo/models/detail/ffe1a0f4e8ba43aeb989251a3f0308e9