qwen-code

Qwen is an open‑source AI framework built for developers who want to run language models inside a terminal. The project ship's the Qwen family of models through GitHub and keeps all weights on Hugging Face and ModelScope, so you can pull them with standard Python libraries.

At the core is Qwen3, a large language model that accepts a plain‑text prompt and streams token output back to the terminal. The model is designed for coding tasks, so typical inputs are code snippets, function signatures or natural‑language descriptions of algorithms. The output is either a code completion, a refactored piece of code, or an explanation of a programming concept. Because the model runs locally, you can experiment with prompt templates or tweak generation settings like temperature and max tokens directly in the command line.

To keep outputs safe, Qwen3Guard is an additional safety filter that sits between the model and the user. It reads every token as it is generated, classifies it into risk levels and categories, and can terminate or flag content that violates predefined rules. Guardship is especially useful when the model is exposed to untrusted prompts. The guardrail works on both prompts and responses, which means you can use it to moderate code that might contain vulnerabilities before it is executed.

For image work, Qwen‑Image is a 20‑billion‑parameter MMDiT model that excels at embedding complex text into images. You give it a source image and a text prompt describing what you want to add or change; the model outputs a new image with the requested text rendered in context, handling multi‑line layouts and paragraph‑level semantics. Qwen‑Image‑Edit builds on that by feeding the same image into Qwen2.5‑VL for visual semantic control and a VAE encoder for appearance control. That allows precise editing of existing text in an image, such as updating a title or correcting a caption while preserving background style.

Translation support comes from Qwen‑MT (qwen‑mt‑turbo). The API accepts a source sentence and a target language code and returns a translated sentence. The model claims coverage for 92 official languages and dialects, aiming to serve users who need quick, multi‑lingual conversion. Because the models are relayed through the same open‑source channel, users can inspect the underlying weights and finetune for specialized domains.

Key features