How to run Qwen 3.8 without a GPU

There are two questions hiding inside "how do I run Qwen 3.8 without a GPU", and they have different answers depending on which Qwen 3.8 you mean.

If you mean the dense 27B, you can run it on your own machine, it is Apache 2.0, and it fits on a single graphics card at 4-bit. We wrote that one up separately in how to run Qwen 3.8 27B on your own computer, and this guide will not repeat it.

If you mean Qwen 3.8 Max or Qwen 3.8 A95B, the flagships, then no consumer hardware runs them and no quantization changes that. A95B is 2.4 trillion parameters in total, and all of them have to sit in memory while it works. That is a rack. The realistic choice is not local against hosted, it is hosted against not using the model.

This guide is about the second case.

Step 1: pick how you want to pay

Every plan and every credit pack carries the whole chat catalog, so both flagships are in the picker whichever you take. Nothing on this page is reserved for a higher tier.

The cheapest way in is a €5 credit pack: 165,000 credits, no subscription, and they do not expire. The Hosted plan is €19 a month with 900,000 credits, and Pro and Max sit above it with 2,350,000 and 5,000,000. What you are choosing is a budget, not a catalog. The pricing page has all of them side by side.

Step 2: buy in

Buy a pack from the credits tab, or start on Hosted if you want a monthly budget. Card details go to Stripe and the payment is secured with 3D Secure.

When the payment lands the Studio opens on your account. Nothing downloads and there is no driver to install.

Step 3: pick Max or A95B

Open the chat tab, click the model picker, and you will find both under their catalog names, Qwen 3.8 Max and Qwen 3.8 A95B.

The short version of the difference: Max is the cheaper of the two per token and it lets you switch the reasoning off. A95B is the open weight 2.4 trillion parameter checkpoint with 95 billion active per token and a 262,144 token context, and it reasons on every single turn whether you want it to or not. The full comparison is here.

Neither of them reads images. If you need to paste a screenshot, that is a different model in the same picker.

Step 4: budget the month before you spend it

Credits translate straight from the catalog price, and one credit is $0.00001.

On Qwen 3.8 Max, a million output tokens draws 495,100 credits. A €5 pack of 165,000 credits spent entirely here is therefore roughly 333,000 output tokens, and a Pro month of 2,350,000 credits is roughly 4.7 million. On A95B a million output tokens draws 600,000 credits, so the same pack is roughly 275,000 and the Pro month roughly 3.9 million.

Both numbers are ceilings that nobody reaches. Input tokens come out of the same wallet, a long conversation resends its whole history every turn, and the same pool pays for images and video. Treat them as the shape of the budget rather than a forecast.

Step 5: use the Think button on Max

On Max the reasoning is genuinely optional, and switching it off is the largest single lever on what a session costs. Leave it on for work that has to be derived. Turn it off for extraction, reformatting, translation and summarising, where you already know the shape of the answer and are only paying for the model to talk itself into it.

On A95B there is no such lever. It reasons every turn by design, so plan for a reasoning pass in every answer and pick it deliberately rather than by default.

One more thing about the effort control: on these two models it is not the knob it is on the GLM 5.3 pair. What actually moves the reasoning on Max is the Think button, not the effort rung.

Step 6: call it from your own code

Go to Settings, then Cloud API keys, and create a key. It starts with lu_ and it is shown once, so paste it into your password manager before you close the dialog. The key spends your credits and cannot read or change your account.

The base URL is https://lu-labs.ai/api/inference/v1 in the OpenAI chat completions shape, so Aider, LibreChat, your own script or plain curl all work unchanged.

curl https://lu-labs.ai/api/inference/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen/Qwen3.8-Max",
       "messages":[{"role":"user","content":"Hello"}]}'

The other id is Qwen/Qwen3.8-2.4T-A95B. Both take tool definitions in the standard tools parameter and call them natively, so agent scaffolding needs no special handling.

Troubleshooting

Neither model is in my picker. Check which account you are signed into, and refresh the picker. Both flagships are in the catalog every plan and every paid pack carries, so a plan is not what is missing. A short picker usually means the demo Studio, which needs no account and does not carry the full list.

An image attachment is rejected. Both of these are text in, text out. Switch to a vision model in the same picker for that turn.

The reply starts with a pause. Cloud runs on shared hardware, so at busy hours a request waits a few seconds in a queue before it starts streaming.

Credits went faster than expected on A95B. It reasons on every turn and there is no way to stop it. If the work does not need that, Max with thinking off is the cheaper answer for the same job.

I wanted the model on my own machine. That is the 27B, and it is a real option: our local guide covers the quantization sizes and the one setting that trips people up.

Where to go next

The model page for both flagships, with the plan and the numbers on one screen, is run Qwen 3.8 without a GPU. For the difference between the two in detail, read Qwen 3.8 Max vs A95B.