How to Run Qwen 3.6 Without a GPU

Qwen 3.6 35B A3B is a mixture of experts model, which means only about 3B parameters are active for any single token. People read that and conclude it must be light. It is not: all 35 billion weights still have to sit in memory while it runs, and the small active count buys you speed, not a smaller download.

That gap between "3B active" and "35B resident" is why so many people install a runner, watch the load fail, and assume they did something wrong. They did not. This guide walks through the check that settles it, and then the route that gets you using the model anyway. The short version of the second half is on the run Qwen 3.6 without a GPU page, and the architecture behind the confusion is in what 3B active parameters actually buys you.

Step 1: do the memory check before you download anything

At the 4-bit quantization almost everyone uses, this model is roughly 18 to 20 GB of weights. That is before the context window takes its share, and the context share grows with every token you feed it.

So the floor is:

• A 24 GB graphics card, and it will be comfortable.
• A Mac with 32 GB of unified memory, and it will work at a slower token rate.
• 8 or 16 GB of anything, and it will either refuse to load or crawl badly enough that you stop using it.

Check your actual number before you spend an evening on this. Our guide on how much RAM you need for local AI walks through where the memory goes and why the context window costs more than people budget for.

If you cleared the bar, run it locally. Genuinely. Offline, free, and nothing leaves your disk. The rest of this guide is for everyone else.

Step 2: get an account, or just credits

Two entry points, and only one of them is a subscription.

Hosted is €19 a month with 900,000 credits. Qwen 3.6 35B A3B is in that catalog, alongside 15 chat models in total.

A credit pack needs no subscription: €5 for 165,000 credits, €10 for 350,000, €25 for 900,000. Any pack is available from your first purchase, and the credits never expire. If you are testing whether this model fits your work, buy the small pack and skip the plan. The full table is on the pricing page.

Start on Hosted if you want the monthly budget. Card details go to Stripe and the payment is secured with 3D Secure.

Step 3: pick the right Qwen in the picker

Open the Studio and click the model picker. Two things to know:

Qwen 3.6 35B A3B is the one in the Hosted catalog. This is the model this guide is about.

Qwen 3.6 27B is a different model and it is Pro and Max only. If you are on the entry plan and cannot find it, that is why, not a bug.

Step 4: use the Think toggle on purpose

Qwen 3.6 35B A3B has a thinking mode you switch on. Reasoning passes cost output tokens, and output tokens cost credits, so the toggle is a spending decision as much as a quality one.

Leave it off for lookups, rewrites, format conversions and anything where you already know what a good answer looks like. Turn it on for multi-step problems where the model getting the middle wrong ruins the end: planning, debugging from a stack trace, anything with arithmetic in it.

The mistake to avoid is leaving it on permanently because it "seems better". On short factual questions it burns credits producing working nobody reads.

Step 5: paste an image

This model accepts image input. Drop a screenshot into the chat and ask about it directly. A prompt that works well:

Here is a screenshot of an error dialog from a piece of software I do not
recognise. Tell me: what program is this most likely from, what the error
means in plain words, and the first thing I should check. Say "unsure" for
any of the three you cannot answer from what is visible.

The "say unsure" clause matters more with vision than with text, because a model reading a blurry screenshot will otherwise fill in the gaps confidently.

Step 6: call it from your own code

Settings, then Cloud API keys, then create a key. It begins with lu_ and is shown once, so save it immediately. The key spends your plan credits and cannot read or change your account.

The base URL is https://lu-labs.ai/api/inference/v1 and the shape is OpenAI chat completions:

curl https://lu-labs.ai/api/inference/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen/Qwen3.6-35B-A3B",
       "messages":[{"role":"user","content":"Convert this CSV header row to a TypeScript interface: id,created_at,owner_email,is_active"}]}'

The model id is spelled exactly as above. Anything that speaks the OpenAI chat format works, including Aider, LibreChat and your own scripts.

Step 7: keep the credits in view

Credits are a single pool for chat, code, image and video. A chat turn is cheap next to a render: Flux Schnell is 300 credits an image, Flux 2 Dev is 1,200, and a 5 s LTX-2 clip is 8,000. One credit is $0.00001, and as a chat reference point Hermes 3 70B in the same catalog costs 0.07 credits per output token, which makes a €5 pack roughly 2.3 million output tokens on it.

The practical consequence: your chat budget is not really threatened by chatting. It is threatened by an afternoon in the image studio. Watch the wallet after a render session, not after a conversation.

Troubleshooting

It loaded locally but generates a word every few seconds. The weights spilled out of VRAM into system RAM. That works and it is slow. Either drop to a smaller quantization, shorten the context, or stop fighting it.

Qwen 3.6 27B is missing from my picker. It is Pro and Max only. The 35B A3B is the Hosted one.

The Think toggle is not there. Check which model is selected. Not every model in the catalog has one. Hermes 3 405B, for example, has no thinking mode at all.

The reply pauses before it starts. Shared hardware. At busy hours a request waits a few seconds in a queue before streaming begins. It resolves on its own.

My API key returns 401. Keys are shown once at creation and cannot be recovered, only revoked and replaced. Also confirm you sent it as Authorization: Bearer lu_....

Image upload does nothing. Confirm the selected model takes image input. Qwen 3.6 35B A3B does; several catalog models are text only, and the picker does not swap models for you when you paste a picture.

If you would rather keep everything on your own disk after all, the desktop app for Windows and Linux is free, runs local models, and needs no account. There is no native Mac build, so on a Mac the browser Studio is the route.