Run Qwen 3.6 without a GPU.

Why a local Qwen 3.6 wants a big card

Qwen 3.6 35B A3B is a mixture of experts model. Only about three billion parameters are active for any single token, which is why it feels quick, but all 35 billion still have to sit in memory while it runs. At the 4-bit quantization almost everyone uses, that is roughly 18 to 20 GB of weights before the context window takes its share. Our RAM guide for local AI works through the sizes in detail.

In practice that means a 24 GB graphics card, or a Mac with 32 GB of unified memory and the patience for a slower token rate. If your machine has 8 or 16 GB, the model will either refuse to load or crawl. Buying hardware for one model is a strange first step, so the cheaper answer is to keep your machine as the screen and put the weights somewhere else.

What you get in the browser

• Qwen 3.6 35B A3B on the Hosted plan. It reads images as well as text, and the Think switch turns its reasoning pass on when a question earns it.
• Qwen 3.6 27B on Pro and Max. The wider catalog adds it next to the other frontier models, with the same vision input and the same switch.
• 15 chat models on the entry plan, all of them able to call tools, so Agent mode and the coding agent run on any of them.
• Image, video and audio studios on the same account and the same credit pool, plus a personal API key for your own scripts.

How to start in three steps

1. Take the Hosted plan at checkout, or buy a credit pack if you would rather not commit to a month. Card details go to Stripe and the payment is secured with 3D Secure.
2. The Studio opens in your browser. Pick Qwen 3.6 35B A3B in the model picker.
3. Type, paste a screenshot, or switch to Code mode and point the agent at a repository.

What it costs

The honest limits

You are sharing hardware with everyone else, so at busy hours a request can wait a few seconds in a queue before the answer starts streaming. Credits are a single pool for text, images and video, and a long agent session spends faster than a chat. The catalog is curated: we host open-weight models we have priced and tested, so you pick from the list rather than uploading your own base model. Account data is hosted in the EU and you can delete it yourself. We never train on your data and we never sell it. And if you would rather keep everything on your own disk, the desktop app for Windows and Linux is free and runs local models with no account at all.

Questions