Qwen3-Coder

What it is

Qwen3-Coder is an open-weight family of code models released by the Qwen team. The first and largest member, Qwen3-Coder-480B-A35B-Instruct, was announced on 22 July 2025 and described by the team as its most agentic code model to date^[1]. The emphasis on "agentic" matters: rather than only completing a snippet of code in an editor, a model of this kind is meant to work inside a loop — reading files, running commands, observing errors and trying again — until a task is finished.

How it is built

The model uses a mixture-of-experts (MoE) architecture, with 480B total parameters of which 35B are active for any given token^[1]. The published model card lists 62 layers and 160 experts, of which 8 are activated per token^[3]. In a mixture-of-experts model, each token is routed to a small subset of specialised sub-networks instead of passing through the whole model; this keeps the amount of computation per token far below what the total parameter count would suggest, while the full set of weights stays available for the knowledge stored in it.

Pre-training used 7.5T tokens with a 70% code ratio^[1]. A detail of the data pipeline is that Qwen2.5-Coder, an earlier model in the same line, was used to clean and rewrite noisy synthetic data before it entered training^[1]. The model has a 256K-token native context window, extendable to 1M through YaRN extrapolation^[1] — YaRN being a technique for stretching the context length of models that use rotary position encodings, so that longer inputs can be handled without retraining from scratch. A long context is what makes repository-scale work feasible: whole files, test output and command history can stay in view at once.

The agentic behaviour was trained rather than merely prompted. The team ran long-horizon agent reinforcement learning across 20,000 independent environments in parallel on Alibaba Cloud^[1].

Reported results

The Qwen team claimed state-of-the-art results among open models on agentic coding, browser use and tool use, comparable to Claude Sonnet 4, and on SWE-bench Verified without test-time scaling^[1]. The "without test-time scaling" qualifier means the scores were obtained without spending extra sampling or search compute at inference time. The Hugging Face model card separately reports 23.9 on Terminal-Bench 2.0 and 38.7 on SWE-Bench Pro^[3].

Tooling and availability

The model was released alongside Qwen Code, an open-source command-line agent adapted from Gemini CLI^[1]. That pairing reflects how coding models are now typically delivered: a model plus a terminal-based harness that lets it act on a working directory. Qwen3-Coder also works with Claude Code and Cline, and the hosted API model is called qwen3-coder-plus^[1]. The Hugging Face model card lists Qwen3-Coder-480B-A35B-Instruct under the Apache-2.0 license^[3], which allows the weights to be downloaded, modified and deployed commercially.

Where it sits

The family expanded after the first release: a smaller Qwen3-Coder-Flash followed in July 2025, and Qwen3-Coder-Next in February 2026^[2]. The pattern — a large flagship, then a lighter variant for cheaper or local use, then a successor generation — is now common across model lines. Qwen3-Coder's significance is that it placed an openly licensed model with a very large context window and explicit agent training into a space that had been dominated by closed hosted systems, and paired it with compatible command-line tooling so that developers could try it in the same workflows.