Codes and reasons like models ten times its size

You give it a task in plain words: fix this failing test, build a small tool, pull the numbers out of this report. Instead of answering once and stopping, it thinks first, acts, looks at what happened, and tries again until the task is finished. Inside a coding assistant it edits files, runs commands, and reads the terminal output like a patient contractor.

That follow-through is the difference. In public performance tests on real software bugs it fixes more of them than systems many times its size, including some well-known closed models. Until now that kind of result came from data-center giants behind an API, not from weights you can download.

It also reads images and video natively, so a screenshot of an error or a photo of a whiteboard can go straight into the conversation. It holds around 260,000 tokens of context, enough for a whole codebase or a stack of long documents in one session. The license is Apache 2.0, which means you can use it in commercial work without asking anyone.

There is no browser demo for this one, so the way in is the model page on Hugging Face. The weights run on the standard local stacks (Transformers, vLLM, SGLang) and comfortably need a GPU with roughly 24 GB of memory; smaller quantized builds from the community are already appearing. For a first test, connect it to your usual coding agent, hand it a genuinely failing test from one of your repos, and watch what happens after its first attempt fails: recovering from feedback instead of looping is exactly what this generation was trained to do. If you have no GPU, Qwen says a hosted version with a bigger context window is coming soon.

The hardware bar is real: without a roughly 24 GB GPU or a quantized build, you are waiting for the hosted version. Its deliberate thinking mode is on by default, so answers arrive slower than a plain chat model until you turn the depth down. And on the longest terminal-driven sessions the biggest closed models still finish more tasks, so give it scoped jobs rather than open-ended marathons.

• Hand it a bug report and let it find the cause, patch it, and rerun the tests.
• Ask questions across a stack of PDFs or an hour-long recording in one session.
• Paste a screenshot of a broken page or an error dialog and ask what went wrong.
• Run a capable code assistant on your own hardware so client code never leaves the building.
• Turn a rough written brief into a small working tool, start to finish.

Try the demo

View model page

Read this issue on neuralexpedition.com →