Alibaba (Qwen team) · 27.8B parameters · Apache 2.0
What GPU do you need to run Qwen3.8 27B?Memory, cards and speed.
Qwen3.8 27B needs about 18 GB of GPU memory (VRAM) at its smallest (4-bit), so a 24 GB card such as an RTX 4090 runs it. The cheapest card on our list that fits, an RTX 3090, costs about $0.155 an hour. At its largest (16-bit) it needs about 57 GB.
GPU memory
Qwen3.8 27B, by version.
Includes about 1.5 GB for the program and a short conversation (about 4,000 tokens). Long conversations need more. Speeds are ceilings: real speeds are usually half to two-thirds of them. How we work it out
| Version | Memory needed | Card size | Cards that fit | From | Top speed |
|---|---|---|---|---|---|
| 4-bit | about 18 GB | 24 GB | RTX 4090, RTX 5090, RTX 3090, RTX 3090 Ti | $0.155/hr on an RTX 3090 | about 107 tokens/s on an RTX 5090 |
| 8-bit | about 31 GB | 32 GB and up | RTX 5090 | $0.524/hr on an RTX 5090 | about 61 tokens/s on an RTX 5090 |
| 16-bit | about 57 GB | None | More than any consumer card holds | — | — |
Qwen's newest open model, from August 2026. It reads images as well as text, and thinks step by step before it answers unless you turn that off. Most of its layers use a lighter kind of attention, so long conversations add less memory than in older models. Apache 2.0.
Run it
Start Qwen3.8 27B in minutes.
On KernelMachine, Rent a Chat model with it: it loads into the chat app by itself once the rental is ready, straight from Ollama to the machine, and is in the app's model menu. On any machine with Ollama, one command downloads and runs it. Ollama on a rented GPU
$ ollama run qwen3.8:27b
# Model files: huggingface.co/Qwen/Qwen3.8-27B
Common questions
Questions, answered plainly.
How much VRAM does Qwen3.8 27B need?
4-bit: about 18 GB; 8-bit: about 31 GB; 16-bit: about 57 GB. That includes about 1.5 GB for the program and a short conversation; longer conversations need more.
Can an RTX 4090 run Qwen3.8 27B?
Yes. Qwen3.8 27B (4-bit) needs about 18 GB, which fits in the RTX 4090's 24 GB.
How fast does Qwen3.8 27B run?
At 4-bit, the ceiling is about 107 tokens/s on an RTX 5090: each word means reading the whole model from memory once. Real speeds are usually half to two-thirds of that.
Can I run Qwen3.8 27B on KernelMachine?
Yes. Rent a Chat model with Qwen3.8 27B, on this page, starts the chat app on the cheapest card with 24 GB or more, and qwen3.8:27b loads into it by itself once it's ready. Already renting a Chat model? Pick Qwen3.8 27B on the rental's page, or download qwen3.8:27b from the chat app's model menu.
Other models
Compare with other chat models.
Preparing an invite-only beta
Get an email when access opens.
No launch date is set. We'll email you once when KernelMachine opens. Nothing else.