Instructions to use bartowski/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 bartowski/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 bartowski/Qwen3.8-27B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use bartowski/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 "bartowski/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": "bartowski/Qwen3.8-27B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Ollama
How to use bartowski/Qwen3.8-27B-GGUF with Ollama:
```
ollama run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M
```
Unsloth Desktop

How to use bartowski/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 bartowski/Qwen3.8-27B-GGUF:Q4_K_M

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": "bartowski/Qwen3.8-27B-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use bartowski/Qwen3.8-27B-GGUF with Docker Model Runner:
```
docker model run hf.co/bartowski/Qwen3.8-27B-GGUF:Q4_K_M
```

Lemonade

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

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.8-27B-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use bartowski/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 bartowski/Qwen3.8-27B-GGUF:Q4_K_M

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 bartowski/Qwen3.8-27B-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use bartowski/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 bartowski/Qwen3.8-27B-GGUF:Q4_K_M

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 "bartowski/Qwen3.8-27B-GGUF:Q4_K_M" \
  --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"

Llamacpp imatrix Quantizations of Qwen3.8-27B by Qwen

Using llama.cpp release b10896 for quantization.

Original model: https://huggingface.co/Qwen/Qwen3.8-27B

Model details:

Parameter count: 28B
Input support: text, image (with mmproj file) - details
Speculative decoding: yes (MTP) - details
imatrix: yes - details
Perplexity/KLD measured: yes - details

How to run

Prompt format

<|im_start|>system
Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.

{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
<think>

Don't know which to choose? Grab Q4_K_M (17.44GB) - usually a good mix of size and performance. Download instructions available here

Updates 11-09-2026

Re-uploaded using my new tensor-layout!

Here's some graphs showing the relative performance between the previous files and the new ones:

You'll note not everything is 100% perfect, but overall everything gets shoved to the ideal quandrant, which is great!

Note that Q2_K_L, Q3_K_XL, and Q5_K_L are no longer being created and so were not updated, they still exist in this repo. If I see a need to introduce them again, I'll consider it, but I think the curve is already pretty nice.

Available files:

Filename	Quant type	File Size	Split	bpw	PPL	Mean KLD	Same top-p	Notes
Qwen3.8-27B-Q8_0.gguf	Q8_0	29.12GB	false	8.38	6.75	0.0009	98.7%
Qwen3.8-27B-Q6_K_L.gguf	Q6_K_L	24.96GB	false	7.19	6.74	0.0024	97.7%	recommended
Qwen3.8-27B-Q6_K.gguf	Q6_K	23.86GB	false	6.87	6.74	0.0030	97.4%	recommended
Qwen3.8-27B-Q6_K_S.gguf	Q6_K_S	22.86GB	false	6.58	6.74	0.0036	97.4%	recommended
Qwen3.8-27B-Q5_K_M.gguf	Q5_K_M	20.92GB	false	6.03	6.75	0.0053	96.9%	recommended
Qwen3.8-27B-Q5_K_S.gguf	Q5_K_S	19.57GB	false	5.63	6.76	0.0060	96.4%	recommended
Qwen3.8-27B-Q4_K_L.gguf	Q4_K_L	18.82GB	false	5.42	6.76	0.0106	95.5%	recommended
Qwen3.8-27B-Q4_1.gguf	Q4_1	17.83GB	false	5.13	6.76	0.0158	94.5%
Qwen3.8-27B-Q4_K_M.gguf	Q4_K_M	17.44GB	false	5.02	6.76	0.0139	95.0%	recommended
Qwen3.8-27B-IQ4_NL.gguf	IQ4_NL	17.44GB	false	5.02	6.76	0.0150	94.8%
Qwen3.8-27B-Q4_K_S.gguf	Q4_K_S	16.36GB	false	4.71	6.77	0.0156	94.8%	recommended
Qwen3.8-27B-Q4_0.gguf	Q4_0	16.35GB	false	4.71	6.86	0.0265	92.8%
Qwen3.8-27B-IQ4_XS.gguf	IQ4_XS	15.48GB	false	4.46	6.78	0.0188	94.1%	recommended
Qwen3.8-27B-IQ3_M.gguf	IQ3_M	14.86GB	false	4.28	6.92	0.0406	91.1%
Qwen3.8-27B-Q3_K_L.gguf	Q3_K_L	14.12GB	false	4.07	6.90	0.0432	90.7%
Qwen3.8-27B-Q3_K_M.gguf	Q3_K_M	13.40GB	false	3.86	7.03	0.0564	89.8%
Qwen3.8-27B-IQ3_XS.gguf	IQ3_XS	12.80GB	false	3.69	7.00	0.0585	89.4%
Qwen3.8-27B-Q3_K_S.gguf	Q3_K_S	12.74GB	false	3.67	7.08	0.0648	89.0%
Qwen3.8-27B-IQ3_XXS.gguf	IQ3_XXS	12.32GB	false	3.55	7.06	0.0739	88.5%
Qwen3.8-27B-Q2_K.gguf	Q2_K	10.82GB	false	3.12	7.44	0.1612	83.5%
Qwen3.8-27B-IQ2_M.gguf	IQ2_M	10.52GB	false	3.03	7.53	0.1494	83.9%
Qwen3.8-27B-IQ2_S.gguf	IQ2_S	9.68GB	false	2.79	7.73	0.2090	81.0%
Qwen3.8-27B-IQ2_XS.gguf	IQ2_XS	9.09GB	false	2.62	8.02	0.2403	79.7%
Qwen3.8-27B-IQ2_XXS.gguf	IQ2_XXS	8.88GB	false	2.56	8.41	0.2834	78.0%

Download a specific file:

hf download bartowski/Qwen3.8-27B-GGUF --include "Qwen3.8-27B-Q4_K_M.gguf" --local-dir ./

Downloading using the Hugging Face CLI

Click to view download instructions

First, make sure you have the Hugging Face CLI installed:

pip install -U "huggingface_hub[cli]"

Download a specific file:

hf download bartowski/Qwen3.8-27B-GGUF --include "Qwen3.8-27B-Q4_K_M.gguf" --local-dir ./

How to run

These quants run with llama.cpp - installable in one line via llama.app:

curl -LsSf https://llama.app/install.sh | sh
llama-server -hf bartowski/Qwen3.8-27B-GGUF:Q4_K_M

llama-server includes a built-in chat web UI, served at http://localhost:8080 by default.

These quants were made with llama.cpp release b10896 - if this model's architecture is newly supported, you'll need that release or newer to run them.

They also work in: LM Studio · koboldcpp · ramalama · Jan AI · Text Generation Web UI · LoLLMs · Atomic Chat

Multimodal

This model supports image input. Alongside the quants, this repo includes the multimodal projector files mmproj-Qwen3.8-27B-f16.gguf and mmproj-Qwen3.8-27B-bf16.gguf, which pair with any quant above.

llama.cpp downloads the mmproj automatically when using -hf as shown above; if you're loading files manually, pass it with --mmproj.

MTP

This model has MTP (Multi-Token Prediction) layers, and they are included in these quants

MTP layers act as a built-in draft model, letting llama.cpp run speculative decoding for faster generation. To use them, add the following flag to your llama.cpp command:

--spec-type draft-mtp

Note: the MTP layers are stored at Q4_0 in the imatrix quants (except for the Q8_0 quant), since imatrix calibration does not exercise them. Q4_0 is chosen for its speed which massively benefits MTP performance.

Per-tensor layouts

Some of these files were built with a layout computed for this model instead of llama.cpp's standard one-size-fits-all rules. A Q4_K_M is still mostly Q4_K; the extra precision goes to the weights this particular model is most sensitive to. The S, M or L in a name says how much of the model stays at the base precision: about 90 % for S, 70 % for M and 50 % for L. An _L name is simply the large size of its family. Q4_K_L is to Q4_K_M what Q4_K_M is to Q4_K_S; Q6_K_S, Q6_K and Q6_K_L are the small, medium and large sizes of Q6_K, with Q6_K_L about halfway to Q8_0. In earlier releases an _L name meant the embedding and output weights were kept at Q8_0; in these files it means the larger size of the base type. There is no size target, so each file's bits per weight is reported rather than promised.

The layout each of these files was built with is published in the layouts/ folder: <file>.tensor-types.txt is the exact --tensor-type-file given to llama-quantize, and <file>.layout.json records how it was computed, including the generator version, the llama.cpp release and the commit, so any of them can be rebuilt.

Checked on this model before any of these files were released: Q6_K reached 0.96×, Q4_K_M 0.94×, Q3_K_M 0.79× and IQ2_XXS 0.78× the KL divergence of the standard layout at the same file size.

These files replace an earlier release of this model that used the standard layout. Every file of both releases was measured the same way, against the same bf16 reference on the same text, and the full side-by-side is in the "Compared with the previous release" fold below.

Layout details

Files built from a computed layout:

Quant	Size	Body bits/weight	File bits/weight	Body kept at base type
Q6_K_L	24.96GB	7.41	7.19	50 %
Q6_K	23.86GB	7.04	6.87	70 %
Q6_K_S	22.86GB	6.71	6.58	90 %
Q5_K_M	20.92GB	6.13	6.03	70 %
Q5_K_S	19.57GB	5.69	5.63	90 %
Q4_K_L	18.82GB	5.49	5.42	50 %
Q4_K_M	17.44GB	5.04	5.02	70 %
IQ4_NL	17.44GB	5.04	5.02	70 %
Q4_K_S	16.36GB	4.69	4.71	90 %
IQ4_XS	15.48GB	4.41	4.46	90 %
IQ3_M	14.86GB	4.25	4.28	50 %
Q3_K_L	14.12GB	4.00	4.07	50 %
Q3_K_M	13.40GB	3.77	3.86	70 %
IQ3_XS	12.80GB	3.57	3.69	90 %
Q3_K_S	12.74GB	3.55	3.67	90 %
IQ3_XXS	12.32GB	3.41	3.55	70 %
Q2_K	10.82GB	2.98	3.12	70 %
IQ2_M	10.52GB	2.88	3.03	70 %
IQ2_S	9.68GB	2.60	2.79	70 %
IQ2_XS	9.09GB	2.40	2.62	90 %
IQ2_XXS	8.88GB	2.34	2.56	70 %

Checked on this model: the computed layout against the standard one, measured by KL divergence against the unquantized model. The ratio compares each computed file with the standard ladder read at that file's own size, so it can differ from the two KLD columns when the two files differ in size.

Quant	Computed layout KLD	Standard layout KLD	Ratio at equal size	Size vs standard file
Q6_K	0.0030 ± 0.0001	0.0031 ± 0.0001	0.96×	+1.7 %
Q4_K_M	0.0139 ± 0.0004	0.0129 ± 0.0002	0.94×	−1.9 %
Q3_K_M	0.0564 ± 0.0009	0.0466 ± 0.0008	0.79×	−8.2 %
IQ2_XXS	0.2834 ± 0.0033	0.3037 ± 0.0034	0.78×	−5.4 %

How it works: the base type is a floor for every body tensor and a fixed share of the body bytes stays at it (90 % for S, 70 % for M, 50 % for L); the remaining bytes go where a cross-model sensitivity prior, measured by KL divergence against the unquantized model, says they buy the most quality. The embedding and output tensors are sized by their share of the file: a small table is kept at Q8_0, a large one follows the file's bitrate. A K-quant and the IQ quant with the same base bitrate (Q3_K_S and IQ3_XS, Q3_K_M and IQ3_S, Q3_K_L and IQ3_M) come out at about the same size; the IQ file is the GPU-oriented twin.

Compared with the previous release

The previous release of this model used llama.cpp's standard layout; these files use the computed one. Both sets were measured against the same bf16 reference on wiki.test.raw over 100 chunks, so the numbers are directly comparable. "Same top-p" is the share of tokens where the quantized model's most likely next token matches the bf16 model's. Q4_0, Q4_1 and Q8_0 are left out because the layout does not touch them. The previous release's Q2_K_L, Q3_K_XL and Q5_K_L have no counterpart here: an _L name now means the large size of its family, so those names are no longer built. Q6_K_S is new in this release, the small size of Q6_K, and has no previous counterpart.

Quant	Previous size	New size	Size change	Previous KLD	New KLD	Previous same top-p	New same top-p
Q6_K_L	24.08GB	24.96GB	+3.7 %	0.0027 ± 0.0001	0.0024 ± 0.0001	97.64 ± 0.10 %	97.70 ± 0.09 %
Q6_K	23.46GB	23.86GB	+1.7 %	0.0031 ± 0.0001	0.0030 ± 0.0001	97.49 ± 0.10 %	97.39 ± 0.10 %
Q6_K_S	—	22.86GB	—	—	0.0036 ± 0.0002	—	97.35 ± 0.10 %
Q5_K_L	21.54GB	—	—	0.0050 ± 0.0002	—	96.92 ± 0.11 %	—
Q5_K_M	20.75GB	20.92GB	+0.8 %	0.0054 ± 0.0001	0.0053 ± 0.0002	96.71 ± 0.11 %	96.87 ± 0.11 %
Q5_K_S	19.68GB	19.57GB	−0.6 %	0.0065 ± 0.0001	0.0060 ± 0.0002	96.25 ± 0.12 %	96.40 ± 0.12 %
Q4_K_L	18.72GB	18.82GB	+0.6 %	0.0124 ± 0.0003	0.0106 ± 0.0002	95.00 ± 0.14 %	95.45 ± 0.13 %
Q4_K_M	17.77GB	17.44GB	−1.9 %	0.0129 ± 0.0002	0.0139 ± 0.0004	95.02 ± 0.14 %	95.00 ± 0.14 %
IQ4_NL	16.33GB	17.44GB	+6.8 %	0.0178 ± 0.0003	0.0150 ± 0.0004	94.22 ± 0.15 %	94.78 ± 0.14 %
Q3_K_XL	16.39GB	—	—	0.0426 ± 0.0007	—	91.01 ± 0.18 %	—
Q4_K_S	16.71GB	16.36GB	−2.1 %	0.0165 ± 0.0003	0.0156 ± 0.0003	94.18 ± 0.15 %	94.77 ± 0.14 %
IQ4_XS	15.57GB	15.48GB	−0.6 %	0.0196 ± 0.0005	0.0188 ± 0.0005	94.03 ± 0.15 %	94.14 ± 0.15 %
IQ3_M	13.90GB	14.86GB	+6.9 %	0.0566 ± 0.0008	0.0406 ± 0.0006	89.66 ± 0.19 %	91.06 ± 0.18 %
Q3_K_L	15.28GB	14.12GB	−7.6 %	0.0437 ± 0.0007	0.0432 ± 0.0007	90.82 ± 0.18 %	90.71 ± 0.18 %
Q3_K_M	14.61GB	13.40GB	−8.2 %	0.0466 ± 0.0008	0.0564 ± 0.0009	90.57 ± 0.18 %	89.81 ± 0.19 %
Q2_K_L	13.08GB	—	—	0.1322 ± 0.0018	—	84.37 ± 0.23 %	—
IQ3_XS	13.33GB	12.80GB	−4.0 %	0.0624 ± 0.0009	0.0585 ± 0.0009	89.18 ± 0.20 %	89.41 ± 0.19 %
Q3_K_S	13.72GB	12.74GB	−7.1 %	0.0686 ± 0.0010	0.0648 ± 0.0010	88.17 ± 0.20 %	88.95 ± 0.20 %
IQ3_XXS	12.63GB	12.32GB	−2.4 %	0.0850 ± 0.0012	0.0739 ± 0.0011	87.34 ± 0.21 %	88.51 ± 0.20 %
Q2_K	11.84GB	10.82GB	−8.6 %	0.1357 ± 0.0019	0.1612 ± 0.0022	83.98 ± 0.23 %	83.53 ± 0.23 %
IQ2_M	10.87GB	10.52GB	−3.2 %	0.1551 ± 0.0021	0.1494 ± 0.0021	83.06 ± 0.23 %	83.91 ± 0.23 %
IQ2_S	10.30GB	9.68GB	−5.9 %	0.2166 ± 0.0026	0.2090 ± 0.0027	79.78 ± 0.25 %	80.98 ± 0.25 %
IQ2_XS	9.99GB	9.09GB	−9.0 %	0.2293 ± 0.0028	0.2403 ± 0.0028	79.53 ± 0.25 %	79.70 ± 0.25 %
IQ2_XXS	9.39GB	8.88GB	−5.4 %	0.3037 ± 0.0034	0.2834 ± 0.0033	76.29 ± 0.27 %	78.04 ± 0.26 %

Perplexity

Perplexity and KL divergence measured against the bf16 model on wiki.test.raw over 100 chunks.

Click to view full KL-divergence statistics

Reference model: bf16
Dataset: wiki.test.raw
Chunks: 100
Base PPL: 6.744 ± 0.10317
llama.cpp build: b10896

Summary

Quant	PPL(Q)	PPL(Q)/PPL(base)	PPL(Q)-PPL(base)	Mean KLD	Mean Δp	RMS Δp	Same top-p
Q8_0	6.75142 ± 0.103316	1.00149 ± 0.000351	0.010043 ± 0.002378	0.000921 ± 6.7e-05	-0.02 ± 0.006%	0.977 ± 0.106%	98.745 ± 0.07%
Q6_K_L	6.73971 ± 0.103022	0.999752 ± 0.000524	-0.001674 ± 0.003535	0.002399 ± 7.4e-05	-0.027 ± 0.009%	1.382 ± 0.058%	97.698 ± 0.094%
Q6_K	6.73999 ± 0.102975	0.999794 ± 0.000568	-0.001387 ± 0.003832	0.002984 ± 0.000106	-0.044 ± 0.01%	1.523 ± 0.066%	97.388 ± 0.1%
Q6_K_S	6.74361 ± 0.103111	1.00033 ± 0.00061	0.002233 ± 0.004113	0.00357 ± 0.000153	-0.031 ± 0.01%	1.599 ± 0.054%	97.353 ± 0.101%
Q5_K_M	6.75391 ± 0.103262	1.00186 ± 0.000762	0.012531 ± 0.00514	0.005287 ± 0.00017	-0.091 ± 0.013%	2.103 ± 0.081%	96.871 ± 0.109%
Q5_K_S	6.76193 ± 0.103379	1.00305 ± 0.000822	0.020546 ± 0.005547	0.006012 ± 0.000195	-0.12 ± 0.014%	2.199 ± 0.069%	96.4 ± 0.117%
Q4_K_L	6.76195 ± 0.103389	1.00305 ± 0.001104	0.02057 ± 0.007443	0.010581 ± 0.000197	-0.099 ± 0.018%	2.841 ± 0.067%	95.451 ± 0.13%
Q4_1	6.76072 ± 0.103288	1.00287 ± 0.001291	0.019337 ± 0.008695	0.015787 ± 0.000324	-0.113 ± 0.022%	3.483 ± 0.085%	94.51 ± 0.143%
Q4_K_M	6.76446 ± 0.103309	1.00342 ± 0.001235	0.023082 ± 0.008317	0.013878 ± 0.000366	-0.153 ± 0.02%	3.254 ± 0.081%	94.996 ± 0.137%
IQ4_NL	6.76445 ± 0.103307	1.00342 ± 0.001293	0.023072 ± 0.008703	0.014962 ± 0.000371	-0.172 ± 0.021%	3.411 ± 0.079%	94.784 ± 0.139%
Q4_K_S	6.77046 ± 0.103353	1.00431 ± 0.001274	0.029081 ± 0.008578	0.015601 ± 0.000295	-0.215 ± 0.022%	3.472 ± 0.08%	94.773 ± 0.139%
Q4_0	6.86425 ± 0.105544	1.01823 ± 0.001735	0.122867 ± 0.011843	0.026524 ± 0.000451	-0.266 ± 0.028%	4.536 ± 0.101%	92.784 ± 0.162%
IQ4_XS	6.78031 ± 0.103638	1.00578 ± 0.001437	0.03893 ± 0.009676	0.018841 ± 0.000471	-0.208 ± 0.024%	3.793 ± 0.079%	94.141 ± 0.147%
IQ3_M	6.92355 ± 0.105518	1.02702 ± 0.002104	0.182168 ± 0.014211	0.040647 ± 0.000622	-0.847 ± 0.036%	5.783 ± 0.11%	91.055 ± 0.179%
Q3_K_L	6.90279 ± 0.106406	1.02394 ± 0.002157	0.161412 ± 0.014737	0.043241 ± 0.000706	-0.355 ± 0.036%	5.685 ± 0.101%	90.714 ± 0.182%
Q3_K_M	7.03035 ± 0.109156	1.04286 ± 0.002543	0.288969 ± 0.01779	0.056425 ± 0.00088	-0.609 ± 0.041%	6.584 ± 0.116%	89.812 ± 0.189%
IQ3_XS	6.99622 ± 0.10673	1.0378 ± 0.002569	0.254844 ± 0.017398	0.058524 ± 0.00086	-1.006 ± 0.042%	6.846 ± 0.114%	89.408 ± 0.193%
Q3_K_S	7.08464 ± 0.110429	1.05092 ± 0.002747	0.343256 ± 0.019408	0.064792 ± 0.000983	-0.6 ± 0.044%	7.094 ± 0.126%	88.953 ± 0.196%
IQ3_XXS	7.05932 ± 0.107866	1.04716 ± 0.002906	0.317942 ± 0.019746	0.07388 ± 0.00114	-1.251 ± 0.048%	7.79 ± 0.134%	88.51 ± 0.2%
Q2_K	7.43849 ± 0.114518	1.10341 ± 0.004572	0.697107 ± 0.031501	0.161216 ± 0.002239	-2.38 ± 0.07%	11.408 ± 0.163%	83.525 ± 0.232%
IQ2_M	7.52818 ± 0.116097	1.11671 ± 0.004474	0.786799 ± 0.031377	0.149377 ± 0.002056	-2.608 ± 0.067%	11.051 ± 0.153%	83.906 ± 0.23%
IQ2_S	7.72954 ± 0.118009	1.14658 ± 0.005361	0.988156 ± 0.036929	0.209045 ± 0.002695	-3.736 ± 0.08%	13.257 ± 0.164%	80.976 ± 0.246%
IQ2_XS	8.025 ± 0.122332	1.19041 ± 0.005935	1.28362 ± 0.04144	0.24027 ± 0.002841	-4.846 ± 0.086%	14.572 ± 0.165%	79.698 ± 0.252%
IQ2_XXS	8.41042 ± 0.130004	1.24758 ± 0.006905	1.66904 ± 0.049626	0.283397 ± 0.003313	-5.486 ± 0.092%	15.735 ± 0.171%	78.039 ± 0.259%

KLD percentile ladder

Quant	Max	99.9%	99%	95%	90%	Median	10%	5%	1%	0.1%	Min
Q8_0	1.0809	0.040699	0.00706	0.002394	0.001448	0.000334	9e-06	2e-06	-1e-06	-8e-06	-3e-05
Q6_K_L	0.973111	0.092973	0.01931	0.006956	0.004484	0.001173	2.3e-05	5e-06	0	-4e-06	-3.2e-05
Q6_K	1.92698	0.105185	0.024555	0.00898	0.005684	0.00142	3.2e-05	7e-06	0	-3e-06	-3.9e-05
Q6_K_S	3.01015	0.139894	0.030559	0.010436	0.006639	0.001579	3.7e-05	9e-06	0	-3e-06	-2e-05
Q5_K_M	2.84051	0.179695	0.051284	0.016652	0.010222	0.002319	5.9e-05	1.4e-05	1e-06	-2e-06	-3.5e-05
Q5_K_S	4.19186	0.204707	0.058225	0.019891	0.011923	0.002638	6.7e-05	1.7e-05	2e-06	-1e-06	-1.3e-05
Q4_K_L	2.23146	0.353113	0.102686	0.036515	0.02187	0.004622	0.000122	3e-05	3e-06	-1e-06	-1.1e-05
Q4_1	4.48169	0.535538	0.15125	0.053299	0.032149	0.006836	0.000176	4.3e-05	5e-06	-0	-8e-06
Q4_K_M	6.50884	0.52521	0.142121	0.045992	0.027477	0.005795	0.000152	3.7e-05	3e-06	-0	-1.1e-05
IQ4_NL	5.86888	0.572585	0.148268	0.049484	0.030007	0.006205	0.00016	4.1e-05	4e-06	-1e-06	-1e-05
Q4_K_S	2.95896	0.534154	0.159605	0.054462	0.032237	0.006696	0.000178	4.5e-05	4e-06	0	-6e-06
Q4_0	4.82004	0.873295	0.263285	0.089089	0.054593	0.011859	0.000302	7.4e-05	8e-06	1e-06	-1e-05
IQ4_XS	7.31304	0.646303	0.189096	0.064107	0.037538	0.007868	0.000206	5.5e-05	5e-06	-0	-2.5e-05
IQ3_M	3.71947	1.19807	0.405264	0.139634	0.084763	0.018335	0.000582	0.00015	1.7e-05	2e-06	-1e-06
Q3_K_L	6.37739	1.27356	0.427362	0.15179	0.090328	0.019294	0.000469	0.000124	1.1e-05	1e-06	-7e-06
Q3_K_M	6.22563	1.74283	0.582466	0.193088	0.116458	0.025449	0.000631	0.000151	1.6e-05	2e-06	-2e-06
IQ3_XS	6.70819	1.54722	0.583696	0.205317	0.12211	0.026238	0.000848	0.000228	2.8e-05	4e-06	-6e-06
Q3_K_S	6.36858	2.01934	0.657329	0.222889	0.135785	0.029351	0.000742	0.000188	1.9e-05	3e-06	-0
IQ3_XXS	6.38787	2.27993	0.733495	0.2525	0.156451	0.033092	0.000938	0.000241	2.6e-05	4e-06	-4e-06
Q2_K	8.55553	4.72795	1.66833	0.554146	0.340936	0.072722	0.00214	0.000494	5.3e-05	8e-06	-0
IQ2_M	11.0817	4.06564	1.48056	0.513973	0.316133	0.068237	0.002261	0.000603	6.7e-05	1.1e-05	-0
IQ2_S	11.4424	5.39947	1.96664	0.727937	0.448283	0.099357	0.003653	0.000949	0.000119	1.9e-05	4e-06
IQ2_XS	10.1463	5.58995	2.10155	0.839563	0.520093	0.123041	0.005352	0.001388	0.000189	2.8e-05	5e-06
IQ2_XXS	11.2416	6.63126	2.46692	0.972586	0.613898	0.147177	0.007022	0.001842	0.000237	3.7e-05	3e-06

Also available as perplexity.md.

imatrix

All quants made using imatrix option with dataset from here. The imatrix is available here: Qwen3.8-27B-imatrix.gguf.

ARM/AVX information

llama.cpp automatically "repacks" weights into an interleaved layout at load time for faster inference on ARM and AVX machines - details in this PR. This once required downloading special Q4_0_4_4/4_8/8_8 files; those are long gone. Online repacking now covers Q4_0, IQ4_NL, and most K-quants, so no special quant choice is needed for CPU inference.

Which file should I choose?

Click here for details

An older (early 2024) but still useful write-up with charts comparing quant performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Hugging Face can also do this math for you: add your hardware in your Local Apps settings and the model page will show which files fit.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.