Alibaba introduces Qwen3-Coder, its most powerful agentic and AI coding model

Alibaba has unveiled a new AI coding model based on Qwen — Qwen3-Coder — described as its most powerful agentic code model so far, and available in several sizes. The flagship version, Qwen3-Coder-480B-A35B-Instruct, features a 480-billion-parameter Mixture-of-Experts architecture with 35 billion active parameters. This model supports a native context length of 256,000 tokens, extendable to 1 million tokens using extrapolation techniques. These technical advances are aimed at improving performance on both code completion and complex agentic tasks.

Alongside this release, Qwen is open-sourcing Qwen Code, a command-line interface tool designed for agentic coding workflows. Based on a fork of Gemini Code, Qwen Code has been modified to feature customized prompts and updated function-calling protocols, optimized to work with Qwen3-Coder. It also integrates with the OpenAI SDK for large language model queries, making it accessible to more developers.

Building on these compatibility improvements, Qwen3-Coder is now usable through Claude Code by obtaining an API key via Alibaba Cloud Model Studio and installing Claude Code. For those preferring direct access, the Qwen3-Coder API is available on the Alibaba Cloud Model Studio platform.