Model guide
Qwen3.8 27B
Qwen3.8 27B is a dense text-and-image model with the same hybrid attention design as Qwen3.6 27B and a 262,144-token context limit. Its Q4 download is about 16.6 GB before the image projector, so plan for a 32 GB computer at a short context and more memory for long conversations. Thinking is on by default; time a short task before judging its speed.
At a glance
- Parameters
- 27B
- Accepts
- Text · Image
- Returns
- Text
- Published context limit
- 262,144 tokensThe model card gives 262,144 tokens natively and says the limit can be extended to 1,000,000. Ollama lists a 256K context window.
- Model format
- safetensors · GGUF
- Software and download hosts
- Ollama · MLX. Follow the exact setup below; this list does not establish support for every format, input type or software version.
- Suggested compression
- Q4_K_M
- Model-weight download
- Q4_K_M weights · 16.56 GBunsloth (community conversion)Qwen3.8-27B-UD-Q4_K_M.gguf
- Image-projector download
- mmproj-BF16.gguf · 0.93 GB. Both downloads total 17.50 GB.
- Languages
- Unknown
- License
- Apache-2.0
Will it fit?
Choose the computer and memory you plan to use. Not sure keeps the answer incomplete; these controls do not detect your hardware.
Enter the positive amount your computer shows. Leave unknown memory as Not sure.
Qwen3.8 27B · Q4_K_M · 4,096 tokens · one session · llama.cpp estimate
Image processing and any companion files need additional memory.
Enter the missing information before choosing a model.
Choose a device and memory type. Enter system RAM or unified memory. Choose a graphics processor, or choose no dedicated GPU if the model will use system RAM.
RAM: not supplied. Estimated need: 22.7 GiB system RAM.
Memory type has not been supplied; choose it before relying on this estimate.
This estimate uses the Q4_K_M model-weight download size, llama.cpp, 4,096 tokens of text context, FP16 conversation memory, one user and a 20% safety allowance. Conversation memory uses the published layer layout; app memory and temporary workspace are allowances. Fit is checked after the operating-system or display reserve. These are estimates, not measured use, and do not establish GPU or architecture support. Image/audio processing, companion files and document indexes need additional memory.
Best for
- Read a short rule set and answer a question that turns on an exception. Check that it applies the visitor rule and the Friday closing time.
Choose something else when
- Allow for 16.6 GB of weights plus the image projector, context and runner before loading. A 16 GB computer cannot hold this model.
Using image input
This model needs an app or library that supports its image-processing components. Follow the exact model’s instructions for preparing image inputs. Some GGUF downloads also need a matching projector file; loading the text model alone does not establish image support.
First setup
Check a representative task
Paste the text below into a new chat with thinking left on, then ask the two questions one at a time.
Office access policy
Normal closing: 18:00
Friday closing: 16:00
Entry: badge or host escort
Visitors: host escort required
A visitor with a badge arrives at 15:30 on Friday. May they enter alone?What time does the office close on Friday?
Expect no (a visitor needs a host escort) and 16:00. An answer that lets the visitor in on the badge has missed the exception. This checks reading a short rule set; it does not check image input or tool use.
Qwen's recommended settings
The model card recommends temperature 1.0, top-p 0.95 and top-k 20 with thinking on, and temperature 0.7, top-p 0.80, top-k 20 and presence penalty 1.5 with thinking off. Thinking is on by default and its text appears inside <think> tags.
This creates a 4,096-token copy of qwen3.8:27b named qwen3-8-27b-short-context. Save the following as a file named Modelfile in your working folder:
FROM qwen3.8:27b
PARAMETER num_ctx 4096
Then run these commands from that folder:
ollama pull qwen3.8:27b
ollama create qwen3-8-27b-short-context -f ./Modelfile
ollama run qwen3-8-27b-short-context
Run the task check. While the model is loaded, run ollama ps in another terminal and check CONTEXT and PROCESSOR.
Ollama lists qwen3.8:27b at 18 GB. It is a separate package from the Hugging Face file used for this memory estimate.
See Ollama context settings and the Modelfile instructions.
Run ollama run qwen3.8:27b (an 18 GB download). On a 32 GB computer, enter /set parameter num_ctx 4096 before the first prompt. Try the task check. Use /bye to leave the chat and ollama stop qwen3.8:27b to unload the model.
Open LM Studio’s Discover tab, search for unsloth/Qwen3.8-27B-GGUF, and download Qwen3.8-27B-UD-Q4_K_M.gguf (16.56 GB). See the download files. Select the file in Chat, set context to 4,096, then load it. These instructions load text only. If it will not load, open LM Studio Model Will Not Load.
Open a terminal in the extracted llama.cpp release folder. On Windows, replace ./llama-server below with .\llama-server.exe:
./llama-server -hf unsloth/Qwen3.8-27B-GGUF -hff Qwen3.8-27B-UD-Q4_K_M.gguf -c 4096 --host 127.0.0.1 --port 8080
The server downloads the selected file. Wait for the model to load, then open the local chat at http://127.0.0.1:8080. This loads text only; image input also needs the projector file. Keep the server terminal open; press Ctrl+C there to stop the server.
Load the model, start a fresh chat and time your usual prompt. Record time until the first useful word and time until the full answer. Note the processor, context and output length when comparing another run. Use the same picture at the same resolution.
- Most of Qwen3.8 27B’s layers (48 of 64) keep a fixed-size memory, so its conversation memory (KV cache) grows slowly with context; only the other 16 layers grow with the conversation.
- The Q4 file used for the estimate is Unsloth’s `UD-Q4_K_M` conversion, not a file published by Qwen. Ollama’s `qwen3.8:27b` package is 18 GB and is packaged separately.
- Q4 can fit some 32 GB computers at a short context; 64 GB is the safer choice for long context or image input.
- Thinking is on by default. It adds waiting time and output tokens before the answer appears.