Disclosure, since half of this post is about a service we run: LU Labs is the hosted layer next to our free open source desktop app for Windows and Linux. The other half is about running Qwen 3.8 on your own hardware, which costs us money rather than making it, and I have written that part first because for a lot of readers it is the right answer.
"How do I run Qwen 3.8 without a GPU" is really two questions wearing one coat, and they have opposite answers depending on which Qwen 3.8 you mean.
First, work out which model you are talking about
Alibaba shipped several things under the same version number in August 2026, and the licences are not the same either.
Qwen 3.8 27B is a dense 27 billion parameter model, Apache 2.0, and it runs on a single graphics card or a decent laptop. This is the one you can download.
Qwen 3.8 Max and Qwen 3.8 A95B are the flagships. A95B is 2.4 trillion parameters in total with 95 billion active per token, and all 2.4 trillion have to be resident while it works. That is a rack, not a desktop. The weights for A95B are public, but under Alibaba's own qwen3.8-max terms rather than Apache, so if you are checking licences for a commercial build, read them rather than assuming.
For the flagships, the choice is not local against hosted. It is hosted against not using the model.
If you meant the 27B: run it yourself
Genuinely, run it. At Q4_K_M the file is about 17 GB, which loads whole onto a 24 GB card with room left for context, spills into system RAM on a 16 GB card (slower, still works), and is comfortable on a 32 GB Apple Silicon Mac. IQ4_XS at about 15.7 GB is the largest quant that stays fully on a 16 GB card.
The one thing that trips almost everyone: Qwen 3.8 ships a chat template, and the template is not decoration. It marks where your message ends and the answer begins, and it carries the reasoning switch. Load a bare GGUF without it and you get one of two failure modes that both look like a bad quantization: it rambles and never stops, or it answers in a clipped voice and forgets the conversation between turns. In llama.cpp the fix is the --jinja flag. LM Studio and Ollama usually bring the template along with the model, so this mostly bites people who pulled a raw file by hand.
The longer version, with the full quant table and the vision file, is in how to run Qwen 3.8 27B on your own computer.
Everything below is about the flagships.
Reaching Max and A95B in a browser or from code
Both are in our chat catalog. Worth stating clearly, because our own older posts said otherwise: the catalog is now the same on every plan and on every credit pack. There is no shortlist, no model reserved for an expensive tier, and a €5 pack bought with no subscription reaches both flagships exactly like a €99 month does. A bigger plan buys credits, not access.
In the browser: sign in at lu-labs.ai, open the chat tab, and both appear under their catalog names, Qwen 3.8 Max and Qwen 3.8 A95B. Nothing downloads and there is no driver to install.
From code: Settings, then Cloud API keys, create one. It starts with lu_ and is shown once, since we keep a hash rather than the key. Base URL https://lu-labs.ai/api/inference/v1, 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 $LU_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"Qwen/Qwen3.8-Max",
"messages":[{"role":"user","content":"Hello"}]}'
import os
from openai import OpenAI
client = OpenAI(
base_url="https://lu-labs.ai/api/inference/v1",
api_key=os.environ["LU_API_KEY"],
)
resp = client.chat.completions.create(
model="Qwen/Qwen3.8-2.4T-A95B",
messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)
Both ids are exact catalog strings. Both take tool definitions in the standard tools parameter and call them natively, so agent scaffolding needs no adapter. Neither reads images: attach a screenshot and the request comes back rejected, which is a property of these two ids rather than a setting you missed. Pick a vision model in the same list for that turn.
What a month of this costs
Credits convert straight from the catalog price, and one credit is $0.00001.
On Qwen 3.8 Max, one million output tokens draws 495,100 credits. On A95B it is 600,000, because A95B is the dearer of the two on both input and output.
| What you buy | Credits | Output tokens on Max | On A95B |
|---|---|---|---|
| €5 pack, no subscription | 165,000 | about 333,000 | about 275,000 |
| Hosted, €19 a month | 900,000 | about 1.8 million | about 1.5 million |
| Pro, €49 a month | 2,350,000 | about 4.7 million | about 3.9 million |
| Max, €99 a month | 5,000,000 | about 10.1 million | about 8.3 million |
All of those are ceilings nobody reaches. Input is billed from the same pool at 0.165 credits per token on Max and 0.2 on A95B, a long conversation resends its whole history every turn, and images and video draw on the same balance. Treat the table as the shape of a budget, not a forecast.
The Think button is the difference between the two
On Max, reasoning is genuinely optional, and switching it off is the single largest 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. There is no parameter that stops it, so plan for a reasoning pass in every single answer and choose the model deliberately rather than by habit.
One more thing about the effort control, since it looks like the same knob and is not: on these two, effort rungs are not what moves the reasoning. On Max it is the Think button. On A95B nothing moves it.
Picking between them
Max is cheaper per token on both sides and it lets you switch the reasoning off, which makes it the default of the two for most work. A95B is the open weight 2.4T checkpoint with 95 billion active per token and a 262,144 token context, and it always reasons, which is either the reason you want it or the reason you do not.
The honest summary: run the 27B at home when the material should not leave your machine, when you want it offline, and when you would rather pay once for a card than per token forever. Reach for a flagship when the problem is genuinely bigger than the hardware in the room. Most people who use both end up doing exactly that.
The full guide with the plan details is at how to run Qwen 3.8 without a GPU.
Top comments (0)