Alibaba launches Qwen3-Coder, a 480B parameter open-source coding AI

Alibaba Cloud has introduced Qwen3-Coder, an open-source large language model (LLM) designed to support code generation, software development tasks, and agentic AI workflows. The release forms part of the company’s Qwen3 series and is available under the Apache 2.0 licence.

The model, named Qwen3-Coder-480B-A35B-Instruct, is a Mixture-of-Experts (MoE) system that comprises 480 billion parameters in total, with 35 billion active parameters at inference time. According to Alibaba, the model is built to facilitate interactive programming and execution in environments that resemble integrated development environments (IDEs).

Supports Agentic coding and long context windows

Qwen3-Coder supports advanced features such as Agentic coding, which allows it to generate, run, and revise code with the help of external tools. The model can operate in a simulated IDE-like environment and has access to a built-in browser for tool usage. It also comes with a command-line interface named Qwen-Code that helps developers interact with the model programmatically.

One of the model’s technical highlights is its context window size of 256,000 tokens, which can be extended up to 1 million tokens through extrapolation methods. This feature is intended to enable the model to handle large-scale codebases and complex software documentation during inference.

Benchmark results and model comparisons

According to performance metrics cited in public model repositories, Qwen3-Coder has demonstrated competitive results on several industry-standard benchmarks. These include the SWE-Bench (Verified) benchmark, which tests AI models on real-world software engineering scenarios.

In these benchmarks, Qwen3-Coder has been reported to perform on par with several proprietary models, such as Anthropic's Claude Sonnet 4 and OpenAI’s GPT-4, in specific coding-related tasks. It also compares against other open-source models like CodeGemma, DeepSeek-V2, and Phind, with Alibaba highlighting favourable results in code completion and reasoning tests.

The company has also released a quantised FP8 version of the model to reduce memory requirements and improve efficiency during deployment.

Availability and strategic context

The Qwen3-Coder model and its associated tools are available for download on GitHub, ModelScope, and Hugging Face, allowing developers and researchers to access the model for experimentation and integration into their workflows.

The Qwen3 family includes a range of models, from 0.5 billion to 235 billion parameters, trained on a corpus of 36 trillion tokens in 119 languages. Alibaba’s open-source approach with Qwen3-Coder follows a broader trend in the AI sector, where large companies are making advanced models available to the wider community for both commercial and non-commercial use.