Alibaba Launches Qwen3-Coder With 480 Billion Parameters and Tools for Autonomous Coding

A coding model built to work through whole tasks

Alibaba Group has unveiled Qwen3-Coder, its most advanced open-source artificial intelligence model for coding to date, with 480 billion parameters, 35 billion active at a time and tools designed to carry out programming tasks across several steps. As of July 24, 2025, the release places Alibaba in closer competition with Chinese rivals and proprietary systems from Anthropic and OpenAI.

The flagship model, Qwen3-Coder-480B-A35B-Instruct, can work with a native context window of 256,000 tokens, extendable to 1 million tokens through extrapolation methods. That gives it room to process large amounts of source code and documentation together, helping it follow connections across a software project.

Alibaba is targeting agentic coding: AI that can plan a task, use development tools, examine the result and revise its work. Instead of stopping at a suggested block of code, the system is designed to take part in a longer workflow, including debugging and changes across a codebase.

The release includes Qwen Code, a command-line tool for giving the AI instructions in everyday language, and an FP8 version intended to reduce the hardware resources needed to run the model. Alibaba has made Qwen3-Coder available under the permissive Apache 2.0 license for commercial and noncommercial use.

Its central pitch combines access with performance. Alibaba says Qwen3-Coder competes with leading proprietary models on coding tasks while allowing developers and businesses greater freedom to deploy and adapt it. The practical question for users is how those capabilities translate into their own projects, computing budgets and requirements for handling data.

How 480 billion parameters become 35 billion active

Qwen3-Coder uses a Mixture-of-Experts architecture, commonly shortened to MoE. Parameters are the learned numerical settings that help an AI model interpret an input and produce an answer. In this architecture, different groups of parameters act as specialists, with only part of the model activated for a particular piece of work.

The flagship contains 480 billion parameters in total, but activates 35 billion at a time. That active portion is about 7.3% of the total, calculated from Alibaba's figures. The design aims to combine the capacity of a very large model with the efficiency of using a smaller portion during processing.

Alibaba presents that balance as useful for businesses running AI at scale. A coding system may need repeated interactions to inspect files, propose a change and assess the outcome. Efficiency therefore matters throughout a task, rather than only when producing the first answer.

The other major technical feature is the context window, which is the amount of information a model can consider together. Tokens are the units into which a model divides text and code. A larger window gives the system more room for instructions, program files, documentation and earlier exchanges.

Qwen3-Coder supports 256,000 tokens natively. Extrapolation methods extend that capacity to 1 million tokens, roughly four times the native window. For a developer, the intended benefit is the ability to bring more of a large project into one working session, including dependencies that connect one part of the software to another.

What went into training the model?

Alibaba pretrained Qwen3-Coder on 7.5 trillion tokens, with code accounting for 70% of that material. The emphasis on programming was intended to build coding skills while retaining general reasoning abilities needed to interpret instructions and work through problems.

Previous Qwen models also helped improve the training data. Alibaba used them to clean and rewrite noisy datasets, aiming to give the new model more reliable material to learn from. Alongside that work, the company optimized the model for long contexts so it could handle large projects and intricate dependencies.

After pretraining, Alibaba used reinforcement learning, a process in which feedback helps a model improve its choices. Its long-horizon approach, called Agent RL, focused on tasks requiring multiple exchanges and tool use rather than a single response.

Alibaba built a training system on Alibaba Cloud capable of running 20,000 independent environments in parallel. Those environments supplied feedback for reinforcement learning at scale, allowing the model to practise working through tasks across a sequence of actions.

The training approach reflects the difference between writing a plausible answer and completing an engineering job. A model handling a longer task must connect earlier decisions with later results. Alibaba's work on data quality, context length and Agent RL was aimed at bringing those abilities together.

From suggesting code to carrying out a workflow

Qwen3-Coder is designed to generate and revise code from high-level instructions, interact with tools such as Git or browsers, plan workflows with several steps, and debug or optimize entire codebases. These are the capabilities Alibaba groups under agentic coding.

A traditional coding assistant may suggest a function or complete a line while a person directs each step. An agentic system takes on more of the sequence: interpreting a goal, choosing a tool, examining what happened and adjusting the next action with less human intervention.

Git, a tool used to manage changes to software, and browser access give such a system ways to interact with a development environment. The model's understanding of context and dependencies helps it connect a requested change with other parts of a project.

For developers, the intended benefit is less time spent on repetitive work and more time for problem solving and system design. The same approach is meant to help with complex tasks that would otherwise require a person to guide a coding assistant through many separate instructions.

Alibaba also designed Qwen3-Coder to work with environments resembling integrated development environments, or IDEs, where programmers edit and manage software. Its tooling includes a built-in browser for tool use, extending the workflow beyond generating text alone.

Qwen Code brings the model into developer tools

Alibaba has released Qwen Code, a command-line interface adapted from the open-source Gemini Code project. A command-line interface, often called a CLI, lets someone work through typed commands in a terminal. Qwen Code allows developers to delegate engineering tasks using natural language.

The tool is optimized for agentic coding and connects the model's capabilities with everyday development workflows. Compatibility with Claude Code is also part of the release, alongside support for popular developer environments and interfaces.

Qwen Code and Qwen3-Coder serve different roles. The model provides the coding and reasoning capabilities, while the interface gives developers a way to direct those capabilities within their work. Together, they are intended to turn a broad request into a sequence of useful engineering actions.

That combination matters to Alibaba's competitive position. A capable model needs a practical route into the places where developers already work. Releasing an accompanying tool gives users a way to try its longer workflows rather than relying only on isolated code generation.

How the benchmark results compare

Alibaba says Qwen3-Coder outperforms domestic competitors including DeepSeek and Moonshot AI's K2, and matches leading US models such as Anthropic's Claude Sonnet 4 and OpenAI's GPT-4 on key coding tasks. Its headline evidence includes results on SWE-Bench Verified, a benchmark that measures how well AI systems fix real software issues.

The cited SWE-Bench results put Qwen3-Coder at 69.6% with 500 turns and at 67.0% in another reported result. The corresponding Claude Sonnet 4 figures are 70.4% and 68.0%. The gaps are 0.8 and 1.0 percentage points respectively, placing Qwen3-Coder close to Claude on those comparisons.

A turn is a step in an interaction. Allowing hundreds of turns gives an agent room to continue working through a problem, which connects the benchmark result with Alibaba's focus on sustained tool use and revision.

Alibaba describes the model as achieving state-of-the-art results and near parity with Claude Sonnet 4 while using fewer active parameters and offering greater efficiency. For organizations evaluating an open-source alternative, the attraction is the combination of competitive coding results, flexibility and potential cost savings.

The numerical comparison here concerns Qwen3-Coder and Claude Sonnet 4. Alibaba's broader performance claim also includes GPT-4, alongside its claims against Chinese competitors. The benchmark figures give readers a concrete measure of the narrow gap with Claude in the results cited at release.

FP8 offers another route to lower deployment costs

Alibaba has also released a quantized FP8 version of Qwen3-Coder. FP8 uses floating point numbers stored in eight bits, reducing the space needed to represent model values. Quantization is the process of representing those values with less numerical precision to reduce resource requirements.

The FP8 version is intended to reduce memory and computing needs, making deployment possible on smaller hardware or allowing more efficient operation in the cloud. That gives businesses another option when balancing model performance against operational costs.

The MoE architecture and FP8 version address efficiency in different ways. MoE limits the portion of the model activated during processing, while FP8 reduces the resources used to represent its values. Both form part of Alibaba's effort to make a large coding model more practical to operate.

Alibaba also plans additional Qwen3-Coder variants that balance performance with lower deployment costs. No release dates were given for those variants or for the company's research into AI agents that can improve themselves.

Open access meets an existing developer audience

Qwen3-Coder is available under Apache 2.0, a permissive license that allows commercial and noncommercial use. Alibaba's decision to release it openly gives developers access to a model they can adapt and deploy, while supporting collaboration and transparency across the programming community.

Developers can access the model and associated tools through the Qwen GitHub repositories, the Qwen Hugging Face collection and Alibaba's Model Studio platform. Those routes offer access alongside the company's own cloud services.

Alibaba has also integrated Qwen3-Coder into Tongyi Lingma, its AI coding assistant. Tongyi Lingma has generated more than 3 billion lines of code since its launch in June 2024, while Qwen-based coding models have recorded more than 20 million downloads globally.

The figures describe different parts of Alibaba's coding ecosystem: generated lines for Tongyi Lingma and downloads for Qwen-based coding models. They show the existing audience into which Alibaba is introducing its latest release.

• June 2024: Tongyi Lingma launched, beginning the period in which Alibaba says the assistant generated more than 3 billion lines of code.
• July 24, 2025: Qwen3-Coder's release brings the flagship model, Qwen Code and an FP8 version into Alibaba's expanding coding ecosystem.

Competition, trust and the choices facing businesses

Alibaba's launch comes amid intensifying competition between Chinese and US technology companies. Restrictions imposed by the United States on AI ecosystem sharing and chip exports form part of the setting in which Chinese companies are investing in domestic AI capabilities.

Those investments include foundational models and advanced GPU hardware. GPUs are processors used for the heavy calculations involved in AI. Building a homegrown combination of models and computing infrastructure is part of the effort by Chinese companies to strengthen their position.

More capable open-source alternatives could reduce AI costs and encourage adoption of Chinese models, particularly in Asia-Pacific markets. For a business considering Qwen3-Coder, that prospect sits alongside the flexibility of its license, its coding performance and the resources needed to run it.

Intellectual property protection, data security and regulatory compliance remain barriers to wider adoption in Western markets. Businesses must consider how a coding system fits their requirements for protecting software and handling information, as well as its technical capabilities.

The rise of sovereign AI adds another consideration. The term describes a preference for AI systems supported locally, especially for applications central to an organization's operations. Trust and national policy may therefore limit the global reach of foreign models even as their performance improves.

What greater autonomy could change

Qwen3-Coder brings together several shifts in software development: larger working contexts, agents that use tools, more efficient model architectures and open access to advanced coding capabilities. Alibaba's release makes those ideas available as a model and accompanying tools rather than as a benchmark result alone.

For enterprises, the intended gains include higher productivity, less reliance on manual coding and automation across more stages of software development. Improving benchmark results from open-source models are giving technology leaders more options when considering innovation and cost savings.

People without programming expertise may also gain more scope to create custom software through natural-language instructions as agentic coding develops. That remains a prospective change, while the immediate release is aimed at helping developers generate, revise and manage code through longer workflows.

Alibaba's next research directions include additional model sizes and agents capable of improving themselves. For now, Qwen3-Coder's concrete contribution is an openly available coding model that pairs a large architecture with selective activation, extended context and tools for carrying out tasks across multiple steps.

Key Points

• Alibaba has released Qwen3-Coder, led by Qwen3-Coder-480B-A35B-Instruct, with 480 billion parameters and 35 billion active at a time.
• The native context window supports 256,000 tokens and can extend to 1 million through extrapolation methods.
• Training used 7.5 trillion tokens, 70% of them code, followed by reinforcement learning supported by 20,000 parallel environments.
• Cited SWE-Bench scores of 69.6% and 67.0% sit close to Claude Sonnet 4's 70.4% and 68.0%.
• Qwen Code supports instructions in natural language, while the FP8 model targets lower memory and computing requirements.
• The Apache 2.0 release permits commercial and noncommercial use, with access through GitHub, Hugging Face and Model Studio.
• Alibaba plans more variants and research into agents that improve themselves, while cost, data protection and national policy remain factors in adoption.