On this page
Yes, at 3-bit. Qwen 3.8 27B’s 3-bit files are 10.9 to 13.1 GB, so they fit on a 16 GB RTX 5080, 5070 Ti or 5060 Ti with room left for your chat. The usual 4-bit file is 16.6 GB, which is more than the card holds, so part of it runs from system RAM and replies slow down. Off Grid AI Desktop runs either one for free on Windows and Linux. You don’t need to buy a 24 GB card to try it.
Download Off Grid AI Desktop (free, Windows, Linux and Mac)
Which Qwen 3.8 27B file fits your card?
Qwen 3.8 27B is a dense 27B model with vision, Apache 2.0 licence and a 262K context (model card). Dense means all 27B parameters work on every token, so the whole file needs to sit on the GPU for full speed.
| File (Unsloth) | Size | 16 GB card | 24 GB card | 32 GB card |
|---|---|---|---|---|
| UD-IQ3_XXS (3-bit) | 10.9 GB | Fits, room for context | Fits | Fits |
| UD-Q3_K_XL (3-bit) | 13.1 GB | Fits, less room | Fits | Fits |
| UD-IQ4_XS (4-bit) | 14.4 GB | Tight, short chats only | Fits | Fits |
| UD-Q4_K_M (4-bit) | 16.6 GB | Spills into RAM | Fits, room for context | Fits |
| Q8_0 (8-bit) | 29.0 GB | Mostly in RAM | Spills into RAM | Fits |
Sizes are from the Unsloth file list, checked 16 November 2026. Vision adds a 0.9 GB file. On Windows the desktop also uses some VRAM, so a 16 GB card has a little less than 16 GB free.
The file isn’t the whole bill. Your conversation needs memory too, and a longer context window needs more. That’s why a 14.4 GB file on a 16 GB card only works with short chats.
Should you buy a bigger card?
Only if 3-bit answers aren’t good enough for your work. A 3-bit file loses some quality against 4-bit, and how much matters depends on the task. Try it first. The step up costs a lot right now.
| Card | VRAM | Launch price (MSRP) | Lowest deal, Nov 2026 | Qwen 3.8 27B |
|---|---|---|---|---|
| RTX 5060 Ti 16 GB | 16 GB | $429 | $790 | 3-bit on the card |
| RTX 5070 Ti | 16 GB | $749 | $1,350 | 3-bit on the card |
| RTX 5080 | 16 GB | $999 | $1,900 | 3-bit on the card |
| RTX 4090 (used) | 24 GB | $1,599 | $3,200 | 4-bit on the card |
| RTX 5090 | 32 GB | $1,999 | $6,099 | 8-bit on the card |
US prices from GPU Poet’s November 2026 tracker. The 5080, 5070 Ti and 5060 Ti all have 16 GB, so for this model they hold the same files. A faster card replies faster, but it doesn’t fit a bigger file.
How to run it on Windows or Linux
- Install Off Grid AI Desktop and open it.
- Go to Settings → GPU performance. In Use your NVIDIA GPU, choose Download, then Restart app when it asks.
- Open Models and search for Qwen 3.8 27B. The app’s catalog download is the 16.6 GB 4-bit file. To stay fully on a 16 GB card, search Hugging Face from the same screen for the Unsloth upload and pick a 3-bit file.
- Start a chat and pick the model. In Model settings, leave the backend on Auto or choose CUDA.
- Send a short prompt, then check the backend’s Now value to see what actually loaded.
If you keep the 4-bit file, lower the GPU layers slider in the app’s performance settings until it loads. The layers you take off the card run from system RAM. It works, and each layer you move makes replies slower. You’ll want 32 GB of system RAM, which is the app’s guidance for this model.
Everything runs on your PC. After the download, it works offline.
What goes wrong, and the fix
| Symptom | Check | Fix |
|---|---|---|
| Out of memory when the model loads | File size against free VRAM | Use a 3-bit file, or lower GPU layers |
| Fast at first, slow in long chats | Context window in Model settings | Lower it, or start a new chat |
| Now shows CPU or Vulkan | Whether the NVIDIA pack installed | Download it in Settings → GPU performance and restart |
| Image questions fail | The 0.9 GB vision file | Leave room for it on the card, or drop to a smaller file |
Need the full-size model?
If you want Qwen 3.8 or DeepSeek V4.1 Flash without squeezing them onto a 16 GB card, Off Grid AI Boost is a flat-price API for open models on our GPUs, with both on the list. It starts at $4.99 a month for the first cohort, with regular prices from $9.99 to $199.99. We don’t store your prompts or train on them. That’s our promise today, and the full policy comes with launch.
A used RTX 4090 starts at $3,200 today. Boost’s lowest plan is $4.99 a month. That’s a price comparison only. Boost’s usage limits aren’t published yet, so we’re not claiming it matches a local card on speed or volume.
Boost starts in January; pre-order to lock in half price. API keys arrive by January 2027.
FAQ
Can an RTX 5080 run Qwen 3.8 27B?
Yes. A 3-bit file of 10.9 to 13.1 GB fits on its 16 GB. The 16.6 GB 4-bit file needs some layers in system RAM.
Is the RTX 5060 Ti 16 GB enough for Qwen 3.8 27B?
Yes, for the same 3-bit files. It holds the same models as an RTX 5080 because it has the same 16 GB. It replies slower.
What about an 8 GB or 12 GB card?
The smallest 2-bit files are 7.2 to 9.9 GB, so a 12 GB card can hold one with a short context. On 8 GB, most of the model runs from system RAM. A smaller model such as Qwen 3.5 9B is a better fit.
How much system RAM do I need?
32 GB is Off Grid AI Desktop’s guidance for Qwen 3.8 27B. You need it most if part of the model runs off the card.
Is Off Grid AI Desktop free?
Yes. It’s free and open source on Windows, Linux and Mac, with no account.
Try a 3-bit Qwen 3.8 27B on the card you already have before you pay $3,200 for a 24 GB one. Download Off Grid AI Desktop.