Can I use Qwen 3.8 without a GPU?

Yes, through a hosted endpoint. Qwen 3.8 Max and Qwen 3.8 A95B are both in the LU Labs cloud catalog, which every plan and every credit pack reaches. Nothing is installed on your machine and no card is involved.

Which plan do I need for Qwen 3.8?

Any of them, and no subscription at all if you buy a credit pack. Every plan and every pack reaches the whole chat catalog, so they differ only in credit budget: a 5 EUR pack is 165,000 credits, the Hosted plan is 19 EUR for 900,000, Pro is 49 EUR for 2,350,000 and Max is 99 EUR for 5,000,000.

What does Qwen 3.8 cost per token?

One credit is $0.00001 of wholesale cost. Qwen 3.8 Max draws 0.165 credits per input token and 0.4951 per output token. Qwen 3.8 A95B draws 0.2 and 0.6. Spent entirely on output tokens, a 5 EUR pack of 165,000 credits is roughly 333,000 tokens on Max or 275,000 on A95B, and the Pro budget is roughly 4.7 million or 3.9 million.

Does hosted Qwen 3.8 take images?

No. Neither Qwen 3.8 Max nor Qwen 3.8 A95B accepts image input, and the request is refused at the API layer. If you need a hosted model that reads pictures, GLM 5.3 Flash and Kimi K3 both do, and both sit in the same picker.

Which Qwen 3.8 can I actually run at home?

The 27B dense checkpoint from the same wave. It is Apache 2.0, it fits on a 16 GB card at IQ4_XS (15.7 GB) or a 24 GB card at Q4_K_M (17.1 GB), it has 262,144 tokens of native context, and a 0.9 GB mmproj file adds image input.

What are the model ids?

Qwen/Qwen3.8-Max and Qwen/Qwen3.8-2.4T-A95B, against base URL https://lu-labs.ai/api/inference/v1 with the OpenAI chat completions shape.

How to run Qwen 3.8 without a GPU

Two different questions hide behind this one. If you mean the big Qwen 3.8 checkpoints, the answer is a hosted endpoint, no subscription required, and the numbers are below. If you mean Qwen 3.8 in general, there is a 27B in the same family that runs on a 16 GB card, and that is a better answer for most people. Both routes are here.

What is actually hosted

The cloud catalog carries two Qwen 3.8 entries, and every plan and every credit pack reaches both of them:

	Qwen 3.8 Max	Qwen 3.8 A95B
Model id	`Qwen/Qwen3.8-Max`	`Qwen/Qwen3.8-2.4T-A95B`
Plan	Every plan and every credit pack	Every plan and every credit pack
Credits per input token	0.165	0.2
Credits per output token	0.4951	0.6
Image input	No	No
Thinking	On by default, can be switched off	On by default, cannot be switched off
Tool calls	Native	Native

One credit is $0.00001 of wholesale inference cost. The difference between those two is smaller than the names suggest and the thinking switch is the part that matters, which is why it has its own page.

Which plan, and what it buys

Every plan and every credit pack reaches the whole chat catalog, so a bigger plan buys a bigger budget and nothing else. A 5 EUR credit pack is 165,000 credits and needs no subscription at all. The Hosted plan is 19 EUR a month for 900,000 credits, Pro is 49 EUR for 2,350,000 and Max is 99 EUR for 5,000,000.

Spent entirely on output tokens, that works out to roughly:

Budget	Qwen 3.8 Max	Qwen 3.8 A95B
5 EUR pack, 165,000 credits	333,000 output tokens	275,000 output tokens
Hosted, 900,000 credits	1.8 million	1.5 million
Pro, 2,350,000 credits	4.7 million	3.9 million
Max, 5,000,000 credits	10.1 million	8.3 million

Those are output tokens only. Input tokens come out of the same wallet, so a long context chat lands below the figure in the row.

What else is in the same picker

The other 36 chat models sit next to the two Qwen 3.8 entries rather than behind a plan, and most of them are far cheaper per token. A few worth knowing by name:

GLM 5.3 Flash, which takes images, reasons on every turn and holds a 1M context, at 0.015 credits per input token and 0.05 per output token
Kimi K3, the 2.8 trillion parameter multimodal reasoner, at 0.285 in and 1.425 out
DeepSeek V3.2 at 0.026 in and 0.038 out, with a think toggle
Gemma 4 26B with image input at 0.007 in and 0.034 out
gpt-oss 120B at 0.0037 in and 0.017 out
Ling 3.0 flash at 0.006 in and 0.018 out

If what you wanted from Qwen 3.8 was a strong reasoner without buying a card, GLM 5.3 Flash gets you most of the way there for a tenth of the token price, and it reads pictures, which neither Qwen 3.8 entry does.

The API route

Sign in at lu-labs.ai and go to Settings > Cloud API keys.
Create a key. It carries an lu_ prefix and is shown once, so copy it before the dialog closes.
Point your client at base URL https://lu-labs.ai/api/inference/v1 with the OpenAI chat completions shape.

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": "Explain this stack trace and name the likely cause."}
    ]
  }'

SillyTavern, Aider, LibreChat and anything else that takes a base URL, a key and a model string work against the same endpoint.

The other answer: run the 27B at home

The same Qwen 3.8 wave shipped a 27B dense checkpoint, and unlike the big two it is Apache 2.0 and it fits on hardware people actually own. Real numbers from our own setup guide:

Quant	Size	Practical home
UD-Q3_K_XL	13.4 GB	16 GB cards
IQ4_XS	15.7 GB	16 GB cards, the last one that fits whole
Q4_K_M	17.1 GB	24 GB cards, the sweet spot
Q6_K	22.9 GB	24 GB, barely, or 32 GB

It is a native vision language model, so a 0.9 GB mmproj file gets you image input the hosted Qwen 3.8 entries do not have. Context is 262,144 native. The full walkthrough with the KV cache math and the chat template trap is in how to run Qwen 3.8 27B locally, and there is a mirror of it at lu-labs.ai.

Our free desktop app for Windows and Linux is one way to run it, and it has a Cloud switch on the same account, so the local 27B and the hosted models sit in one picker. That is the setup we actually use: the 27B for the everyday work that stays on the machine, the switch for the prompts that outgrow it.