Deploy Alibaba’s Qwen3.5‑397B‑A17B Model in One Click with PAI‑Model Gallery

Model Overview

Qwen3.5‑397B‑A17B is an open‑source large language model released by Alibaba. It has 397 billion total parameters but only 17 billion active parameters. Compared with the 1‑trillion‑parameter Qwen3‑Max, it achieves higher accuracy on a range of benchmarks while reducing GPU memory consumption by about 60 % and increasing inference throughput up to 19×.

Architecture

The model uses a hybrid design that combines linear attention implemented as Gated Delta Networks with a sparse mixture‑of‑experts (MoE) layer. This yields fast inference and lower cost while preserving capability. Language support has been expanded to 201 languages and dialects. Benchmarks show strong performance on reasoning, code generation, agent tasks, and multimodal understanding.

PAI‑Model Gallery

PAI‑Model Gallery is a component of Alibaba Cloud’s PAI platform that aggregates open‑source models (LLM, CV, NLP, etc.) and provides zero‑code pipelines for training, deployment and inference. The Qwen3.5‑397B‑A17B model is listed there with enterprise‑grade security and automatic cloud‑resource adaptation.

One‑Click Deployment Procedure

Open the model page in PAI‑Model Gallery:

https://pai.console.aliyun.com/#/quick-start/models/Qwen3.5-397B-A17B/intro

.

Click Deploy . The platform offers high‑performance back‑ends based on SGLang and vLLM. Select the desired compute resources (GPU type, number of instances, etc.).

After deployment finishes, retrieve the service endpoint and authentication token from the “View Call Info” button on the service page.

Invoke the model via HTTP API, SDKs, or the PAI online debugging console. Detailed API usage is documented at

https://help.aliyun.com/zh/pai/getting-started/model-gallery-quick-st#art-091a6b66e9v4x

.

Additional Model Support in PAI‑Model Gallery

The gallery continuously adds popular open‑source models such as Qwen, Wan, DeepSeek, Kimi, MiniMax, and PAI‑optimized variants (e.g., Qwen3‑235B‑A22B‑PAI‑optimized) with built‑in EP+PD deployment templates for improved performance.