AI News

A Comprehensive Technical Breakdown of Qwen: Model Matrix, MoE Principles, API Practice, and Enterprise Selection Guide

Alibaba Cloud Developer Community AIUpdated Nov 16 17:49

AI summary

Qwen's flagship Qwen3.8-Max-Preview adopts a MoE architecture, with 2.4 trillion total parameters; each inference activates only 8 of 128 experts, about 35 billion parameters, and the company claims inference speed is improved by more than 5 times. The model family covers Qwen-Max, Plus, Flash, Coder, VL, Omni, and Qwen-Long, among which Max and Plus support a million-token context, and Qwen-Coder's HumanEval score is above 85. The API is compatible with the OpenAI interface, new users can get 70 million free Tokens, and the Token Plan personal edition costs 39 yuan/month.

Body

Only a summary is authorized here. Continue at the validated source link.

Source

Alibaba Cloud Developer Community AI