
Alibaba has unveiled its latest wave of Qwen3 models, and the headline act is Qwen3-Coder, its most advanced agentic AI coding model to date. It is open source under the Apache 2.0 licence, giving developers the freedom to download, run, inspect and adapt it for their own work.
The flagship variant, Qwen3-Coder-480B-A35B-Instruct, has 480 billion parameters in total and uses a mixture-of-experts design, with only 35 billion parameters active for each token. It is available on Hugging Face, GitHub and ModelScope, alongside a quantised FP8 version that makes it easier to run on more modest hardware.
Agentic coding tools are fast becoming the default way software gets written. Most of the strongest options, however, are closed subscription services. An open-weight model that can compete on real engineering benchmarks therefore changes who can build with this kind of technology, and on what terms.
What Qwen3-Coder Actually Does
Agentic coding means more than autocomplete. Qwen3-Coder can generate code, run it, inspect the results, fix errors and work through complex pipelines across an entire codebase, rather than stopping at a single suggestion.
It is built around three capabilities: agentic coding, agentic browser-use and agentic tool-use. Browser-use gives it a built-in browser to look up documentation and live information, while tool-use lets it call external tools and operate in a simulated IDE-like environment.
That browser access matters because it lets the model ground its work in up-to-date information, rather than relying only on what it learned during training.
A Mixture-of-Experts Design That Keeps Costs Down
Qwen3-Coder uses a mixture-of-experts architecture: 480 billion parameters in total, with only 35 billion active per token. Routing sends each piece of work to the expert sub-networks best suited to it, so most of the model stays idle for any given token.
As a result, inference costs are much closer to those of a far smaller model, even though the total capacity is far larger.
Alibaba reports competitive results against leading state-of-the-art models on agentic coding, browser interaction and tool-use benchmarks. Among open-source models, it is state-of-the-art on SWE-Bench Verified, a benchmark built from real-world software issues taken from actual repositories.

A 256K Context Window That Stretches to 1 Million Tokens
Qwen3-Coder has a native 256,000-token context window, which can be extended to 1 million tokens using YaRN. That is roughly enough to hold a substantial codebase in a single session, keeping more files, history and context in view without constant re-prompting.
The scale of the wider Qwen family is considerable: it spans 0.5 billion to 235 billion parameters, was trained on 36 trillion tokens in 119 languages, and Qwen-based coding models have passed 20 million downloads globally.
Qwen3-Coder was released alongside Qwen3-235B, focused on reasoning, and Qwen-MT, for translation. It also follows Alibaba’s Qwen4-Max flagship, which arrived earlier in November with a 1M-token context of its own.
Qwen Code: A Command-Line Coding Agent
Alibaba is also open-sourcing Qwen Code, a command-line tool that lets developers delegate engineering tasks in natural language. It plans the work, edits files, runs commands and reports back when it is done.
Interest has been swift: the qwen-code repository passed 30,000 GitHub stars in mid-November, as open-source coding agents are gaining traction among developers looking for tools they can run and adapt themselves.
Under the hood, it is a TypeScript/Node CLI wrapping the Qwen API in an agent loop with file editing, shell execution and search. It can also point at OpenAI-compatible endpoints, so it is not tied to a single provider.
That design meets developers where they already work, in the terminal, next to git and the test runner. For a team, it means a capable coding agent can run entirely on its own infrastructure.

How It Stacks Up Against the Competition
The competition is moving quickly. Rivals include Microsoft’s MAI Code 2, GitHub Copilot’s expanding agent capabilities with new MCP connectors, and OpenAI’s GPT-6.2 models.
Qwen3-Coder’s pitch is openness: downloadable weights, a permissive licence and an FP8 version for modest hardware. Independent benchmark trackers put its SWE-Bench Verified score at 69.6%.
Closed frontier models still lead on the hardest multi-file refactoring and long-horizon tasks, as noted in Gadgets 360‘s coverage of the release.
There is still a trade-off. Self-hosting the full 480B model still needs substantial GPU capacity, so most teams will run the smaller 30B weights locally and call the API for harder tasks.
Who Should Try Qwen3-Coder
Qwen3-Coder is best for developers and teams who want an agentic coding assistant they can control: self-hosting for privacy, working in regulated industries, or simply using open tools they can inspect and modify.
Tongyi Lingma, Alibaba’s Qwen-powered coding assistant, is set to be upgraded with Qwen3-Coder’s capabilities.
Start small: point Qwen Code at a side project, give it a well-scoped task, and review everything before merging. Agentic tools work best with a human in the loop.
The Bottom Line
Qwen3-Coder is the open-source answer to the agentic coding wave: a 480B-parameter mixture-of-experts model, with 35B parameters active per token, up to 1M tokens of context, the ability to browse the web and use tools, and a lead among open models on SWE-Bench Verified.
Paired with the open-source Qwen Code CLI, it puts a capable coding agent within reach of any developer with the hardware to run it. Closed models still set the pace on the hardest tasks, but the gap is closing, and this challenger is free, open and yours to run.

