News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnQwen3-Coder

Qwen3-Coder

Agentic coding model with 480 billion parameters

Qwen3-Coder is a large language model for software engineering tasks, released by the Qwen team in July 2025. It uses a mixture of experts architecture to handle complex coding and agent workflows, supporting very long contexts for repository analysis.

Share on WhatsApp

  • 480 billion parameters with 35 billion active
  • Native 256K context window extendable to 1M
  • Trained on 7.5 trillion tokens with 70% code
  • Supports agentic coding and tool use
  • Open weights available for research and use

What it is

Qwen3-Coder is a specialised language model designed for software development and agentic tasks. The Qwen Team introduced it in July 2025 as their most capable coding model. The initial release features a variant with 480 billion total parameters, though only 35 billion are active during inference due to its mixture of experts design. This structure allows it to process vast amounts of code while maintaining efficiency. It supports native context windows of 256,000 tokens, which can be extended to one million tokens using extrapolation techniques. This capacity enables the model to understand entire code repositories and manage long-term planning in software projects.

How it came about

The model builds upon previous iterations in the Qwen series, specifically leveraging Qwen2.5-Coder for data cleaning and rewriting. The developers focused on scaling pretraining data to 7.5 trillion tokens, with seventy per cent consisting of code. This approach aimed to enhance coding proficiency while retaining general mathematical and logical abilities. The team also emphasised reinforcement learning from human feedback and automated verification processes to improve performance on real-world engineering benchmarks.

Inside the model

Qwen3-Coder employs a mixture of experts architecture, activating eight experts per token. It uses group query attention with 96 heads for queries and eight for key-value pairs. Training involved extensive reinforcement learning on diverse coding tasks, including multi-turn interactions with development environments. The team utilised a scalable infrastructure running twenty thousand parallel environments to facilitate this agent-based learning. Post-training adjustments focused on long-horizon tasks, allowing the model to plan, use tools, and adapt based on environmental feedback without requiring test-time scaling.

What it can do

The model demonstrates strong performance in agentic coding, browser use, and tool utilisation. Benchmarks indicate results comparable to Claude Sonnet 4 in these areas. It achieves state-of-the-art scores among open models on SWE-Bench Verified. The system is optimised for repository-scale understanding and dynamic data handling, such as processing pull requests. It supports function calling formats compatible with various developer tools, enabling seamless integration into existing workflows.

How to use it

Weights for the Qwen3-Coder-480B-A35B-Instruct variant are available openly. Users can access the model via APIs through Alibaba Cloud Model Studio or DashScope. The team provides a command-line interface called Qwen Code, adapted from Gemini CLI, to facilitate local interaction. Integration with other tools like Cline and Claude Code is supported through custom configurations and proxy APIs.

Questions people ask about Qwen3-Coder

What is the parameter size of Qwen3-Coder?

The model has 480 billion total parameters, but uses a mixture of experts design that activates only 35 billion parameters per token. This allows for high capacity while keeping inference costs manageable.

How long is the context window?

It natively supports 256,000 tokens. Using YaRN extrapolation methods, this can be extended up to one million tokens, making it suitable for analysing large codebases.

Is Qwen3-Coder open source?

Yes, the weights are openly available. The model can be downloaded for local use or accessed via API through Alibaba Cloud services.

What tools does it work with?

It integrates with Qwen Code, Cline, and Claude Code. It uses standard function calling formats compatible with many developer environments and supports OpenAI-compatible API endpoints.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Qwen3-Coder yet, so the facts on this page come from the sources listed above: huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct, qwenlm.github.io/blog/qwen3-coder/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.