Instructions to use KarLexAI/Qwen3.8-27B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use KarLexAI/Qwen3.8-27B-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S
# Run inference directly in the terminal:
llama cli -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S
# Run inference directly in the terminal:
llama cli -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S
# Run inference directly in the terminal:
./build/bin/llama cli -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Use Docker

docker model run hf.co/KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Llama new
LM Studio
Jan

vLLM

How to use KarLexAI/Qwen3.8-27B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "KarLexAI/Qwen3.8-27B-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "KarLexAI/Qwen3.8-27B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Ollama
How to use KarLexAI/Qwen3.8-27B-GGUF with Ollama:
```
ollama run hf.co/KarLexAI/Qwen3.8-27B-GGUF:IQ1_S
```
Unsloth Desktop

How to use KarLexAI/Qwen3.8-27B-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "KarLexAI/Qwen3.8-27B-GGUF:IQ1_S"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use KarLexAI/Qwen3.8-27B-GGUF with Docker Model Runner:
```
docker model run hf.co/KarLexAI/Qwen3.8-27B-GGUF:IQ1_S
```

Lemonade

How to use KarLexAI/Qwen3.8-27B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Run and chat with the model

lemonade run user.Qwen3.8-27B-GGUF-IQ1_S

List all available models

lemonade list

Hermes Agent

How to use KarLexAI/Qwen3.8-27B-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use KarLexAI/Qwen3.8-27B-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf KarLexAI/Qwen3.8-27B-GGUF:IQ1_S

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "KarLexAI/Qwen3.8-27B-GGUF:IQ1_S" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Qwen3.8-27B · KLX GGUF

27B reasoning from 6.2 GB. It still thinks.

The smallest GGUFs of a 27B reasoning model usually stop reasoning: they loop, run out of tokens and miss easy problems. KLX files keep more of the model at every size. At 6.2 GB, the KLX file solves 53.8% of AIME 2025+2026; Unsloth's file of the same size solves 1.3%. Standard GGUF, no fine-tuning, no custom runtime.

✨ Highlights

🧠 53.8% on AIME 2025+2026 and 87.8% on MATH-500 from a 6.19 GB file. Unsloth's 6.19 GB file scores 1.3% and 32.6%.
🎯 Near full quality from 10.9 GB. 94.2% AIME, 94.3% IFEval and 99.0% MATH-500, against 96.0%, 93.4% and 99.2% for full BF16 (about 54 GB).
🔌 Plain GGUF. Only standard llama.cpp tensor types: nothing to patch, no fork to build.

📦 Pick your file

File	Size	AIME 2025+2026	IFEval	MATH-500	Pick it for
Qwen3.8-27B-KLX-IQ1_S.gguf	6.19 GB	53.8	50.5	87.8	the smallest download
Qwen3.8-27B-KLX-IQ2_XXS.gguf	7.21 GB	82.7	88.5	98.4	the biggest step up per GB
Qwen3.8-27B-KLX-IQ2_XS.gguf	8.00 GB	89.4	93.5	97.8	strong reasoning at 8 GB
Qwen3.8-27B-KLX-IQ3_XXS.gguf	10.93 GB	94.2	94.3	99.0	near full quality
Qwen3.8-27B-KLX-IQ4_XS.gguf	14.25 GB	95.2	94.3	98.8	the closest to full quality
Qwen3.8-27B BF16, for reference	about 54 GB	96.0	93.4	99.2

The file names give the size class. Each file mixes several standard tensor types.

🥊 Head to head

Same harness, same questions, paired differences in points with one standard error.

Size	KLX file	Against	AIME 2025+2026	IFEval	MATH-500
6.2 GB	IQ1_S	Unsloth UD-IQ1_S (6.19 GB)	+52.5 ± 4.9	+28.8 ± 2.6	+55.2 ± 2.5
7.2 GB	IQ2_XXS	Unsloth UD-IQ2_XXS, 19 Aug file (7.27 GB)	+8.75 ± 2.6	+8.3 ± 1.9	+1.8 ± 0.9
8 GB	IQ2_XS	ISTA GSQ-RCO IQ2_XS (8.42 GB)	+0.2 ± 1.5	-1.1 ± 1.1	+0.2 ± 0.8
8 GB	IQ2_XS	Unsloth UD-IQ2_S, 19 Aug file (8.37 GB)	+1.0 ± 1.4	+2.8 ± 1.3	-0.8 ± 0.8
10.9 GB	IQ3_XXS	Unsloth UD-IQ3_XXS (10.93 GB)	+1.0 ± 1.2	-0.4 ± 0.8	+0.2 ± 0.4
14.3 GB	IQ4_XS	Unsloth UD-IQ4_XS, 19 Aug file (14.25 GB)	-0.6 ± 0.9	-1.1 ± 0.8	+0.2 ± 0.5

From 10.9 GB up, both makers' files are close to full quality, and the benchmarks above cannot separate them.

Note, 12 Nov 2026: on 10–11 Nov Unsloth replaced UD-IQ2_XXS, UD-IQ2_S and UD-IQ4_XS with new "Dynamic v3.1" files. The comparisons with those three files were measured against Unsloth's earlier files of 19 Aug 2026 and have not yet been repeated with the new ones. UD-IQ1_S and UD-IQ3_XXS are unchanged, so those comparisons are current. Setup, every score and the file hashes: BENCHMARKS.md.

🚀 Run it with two commands

hf download KarLexAI/Qwen3.8-27B-GGUF Qwen3.8-27B-KLX-IQ1_S.gguf --local-dir .
llama-server -m Qwen3.8-27B-KLX-IQ1_S.gguf -ngl 99 -c 32768 --jinja

Then open http://localhost:8080 and chat. Thinking is on by default, and Qwen's sampling settings (temperature 1.0, top-k 20, top-p 0.95) are stored in the file.

🖼️ Images

Qwen3.8 also reads images. Download the vision projector (about 0.9 GB) and add --mmproj:

hf download KarLexAI/Qwen3.8-27B-GGUF mmproj-F16.gguf --local-dir .
llama-server -m Qwen3.8-27B-KLX-IQ1_S.gguf --mmproj mmproj-F16.gguf -ngl 99 -c 32768 --jinja

mmproj-BF16.gguf works the same way, with any of the files above (we tested IQ1_S and IQ4_XS on llama.cpp commit 351025ec).

🧾 The fine print

Post-training quantization of Qwen/Qwen3.8-27B. No fine-tuning. The architecture, tokenizer and chat template are unchanged; the multi-token prediction (MTP) head is not included. The vision encoder is in the separate mmproj files, unquantized (F16 and BF16), converted with upstream llama.cpp.
Every number above comes from one harness on stock llama.cpp, in thinking mode with Qwen's sampler.
Apache-2.0, like the base model. See NOTICE.

Made by KarLex AI. If these files save you memory, a ❤️ on this repo helps others find them.

_{Not affiliated with Qwen, Unsloth or ISTA.}

Downloads last month: 1,944

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

1-bit

2-bit

3-bit

4-bit

Model tree for KarLexAI/Qwen3.8-27B-GGUF

Base model

Qwen/Qwen3.8-27B

Quantized

(1,793)

this model