ICONNAI_ICONN-1-GGUF Open-Source Model - Multiple quantization options to meet different performance and quality requirements

ICONNAI ICONN 1 GGUF

Developed by bartowski

Quantized version of ICONN-1, offering multiple quantization options to meet different performance and quality requirements

Large Language Model Open Source License:Other #Multiple quantization versions #High-performance inference #Large model optimization

Downloads 609

Release Time : 6/19/2025

Model Overview

This project provides the Llamacpp imatrix quantized version of ICONNAI's ICONN-1 model, including multiple quantization options from Q2 to Q8, suitable for different hardware configurations and usage scenarios

Model Features

Diverse quantization options

Provide more than 20 quantization versions from Q2 to Q8 to meet different hardware configurations and performance requirements

imatrix quantization technology

Use the imatrix option of llama.cpp for quantization to optimize model performance

Hardware adaptation optimization

Support online weight repacking for ARM and AVX CPUs to improve the running efficiency on specific hardware

High-quality recommended options

Provide multiple high-quality quantization versions such as Q6_K_L and Q5_K_M as recommended choices

Model Capabilities

Text generation

Dialogue system

Instruction understanding

Use Cases

Local deployment

Run in LM Studio

The quantized model can be directly run in LM Studio

Integrate with llama.cpp

Can be integrated with any project based on llama.cpp

Research application

Research on quantization technology

Can be used to study the impact of different quantization methods on model performance

🚀 Llamacpp imatrix Quantizations of ICONN-1 by ICONNAI

This project provides quantized versions of the ICONN-1 model by ICONNAI, offering various options for different performance and quality requirements.

🚀 Quick Start

ACCESS REQUESTS ENABLED

Be aware that there is heavily negative feedback. Instead of deleting or privating the model, it's gated to prevent unawareness during downloading: Check the discussions

I won't read any info you submit to gain access; it will be automatically approved. This is just an extra "I understand" step.

✨ Features

Quantized Models: Offers multiple quantized versions of the ICONN-1 model.
Multiple Running Options: Can be run in LM Studio or directly with llama.cpp and other llama.cpp-based projects.
Different Prompt Formats: Uses a specific prompt format for interaction.

📦 Installation

Using huggingface-cli

Click to view download instructions

First, ensure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, target the specific file you want:

huggingface-cli download bartowski/ICONNAI_ICONN-1-GGUF --include "ICONNAI_ICONN-1-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it's split into multiple files. To download them all to a local folder, run:

huggingface-cli download bartowski/ICONNAI_ICONN-1-GGUF --include "ICONNAI_ICONN-1-Q8_0/*" --local-dir ./

You can specify a new local-dir (ICONNAI_ICONN-1-Q8_0) or download them all in place (./).

💻 Usage Examples

Prompt Format

No chat template is specified, so the default is used. This might be incorrect; check the original model card for details.

<s>[SYSTEM_PROMPT]{system_prompt}[/SYSTEM_PROMPT][INST]{prompt}[/INST]

📚 Documentation

Model Information

Property	Details
Quantized By	bartowski
Pipeline Tag	text-generation
License Link	LICENSE
License Name	iconn
Base Model	ICONNAI/ICONN-1
License	other
Base Model Relation	quantized
Tags	emotional-ai, ICONN, chatbot, base
CO2 Eq Emissions	Emissions: 1.34 Source: CodeCarbon Training Type: pretraining Geographical Location: US-West Hardware Used: 9 x B200
Extra Gated Prompt	Be aware there are reported issues, read up before wasting bandwidth

Download a File

Filename	Quant type	File Size	Split	Description
ICONN-1-Q8_0.gguf	Q8_0	89.23GB	true	Extremely high quality, generally unneeded but max available quant.
ICONN-1-Q6_K_L.gguf	Q6_K_L	69.22GB	true	Uses Q8_0 for embed and output weights. Very high quality, near perfect, recommended.
ICONN-1-Q6_K.gguf	Q6_K	68.89GB	true	Very high quality, near perfect, recommended.
ICONN-1-Q5_K_L.gguf	Q5_K_L	60.04GB	true	Uses Q8_0 for embed and output weights. High quality, recommended.
ICONN-1-Q5_K_M.gguf	Q5_K_M	59.63GB	true	High quality, recommended.
ICONN-1-Q5_K_S.gguf	Q5_K_S	57.83GB	true	High quality, recommended.
ICONN-1-Q4_1.gguf	Q4_1	52.63GB	true	Legacy format, similar performance to Q4_K_S but with improved tokens/watt on Apple silicon.
ICONN-1-Q4_K_L.gguf	Q4_K_L	51.40GB	true	Uses Q8_0 for embed and output weights. Good quality, recommended.
ICONN-1-Q4_K_M.gguf	Q4_K_M	50.91GB	true	Good quality, default size for most use cases, recommended.
ICONN-1-Q4_K_S.gguf	Q4_K_S	47.84GB	false	Slightly lower quality with more space savings, recommended.
ICONN-1-Q4_0.gguf	Q4_0	47.63GB	false	Legacy format, offers online repacking for ARM and AVX CPU inference.
ICONN-1-IQ4_NL.gguf	IQ4_NL	47.45GB	false	Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference.
ICONN-1-IQ4_XS.gguf	IQ4_XS	44.85GB	false	Decent quality, smaller than Q4_K_S with similar performance, recommended.
ICONN-1-Q3_K_XL.gguf	Q3_K_XL	44.13GB	false	Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability.
ICONN-1-Q3_K_L.gguf	Q3_K_L	43.55GB	false	Lower quality but usable, good for low RAM availability.
ICONN-1-Q3_K_M.gguf	Q3_K_M	40.23GB	false	Low quality.
ICONN-1-IQ3_M.gguf	IQ3_M	36.94GB	false	Medium-low quality, new method with decent performance comparable to Q3_K_M.
ICONN-1-Q3_K_S.gguf	Q3_K_S	36.36GB	false	Low quality, not recommended.
ICONN-1-IQ3_XS.gguf	IQ3_XS	34.45GB	false	Lower quality, new method with decent performance, slightly better than Q3_K_S.
ICONN-1-IQ3_XXS.gguf	IQ3_XXS	32.40GB	false	Lower quality, new method with decent performance, comparable to Q3 quants.
ICONN-1-Q2_K_L.gguf	Q2_K_L	31.41GB	false	Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable.
ICONN-1-Q2_K.gguf	Q2_K	30.76GB	false	Very low quality but surprisingly usable.
ICONN-1-IQ2_M.gguf	IQ2_M	27.74GB	false	Relatively low quality, uses SOTA techniques to be surprisingly usable.
ICONN-1-IQ2_S.gguf	IQ2_S	25.29GB	false	Low quality, uses SOTA techniques to be usable.
ICONN-1-IQ2_XS.gguf	IQ2_XS	24.77GB	false	Low quality, uses SOTA techniques to be usable.
ICONN-1-IQ2_XXS.gguf	IQ2_XXS	22.30GB	false	Very low quality, uses SOTA techniques to be usable.

Embed/Output Weights

Some of these quants (Q3_K_XL, Q4_K_L etc) use the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of the normal default.

ARM/AVX Information

Previously, you'd download Q4_0_4_4/4_8/8_8, and their weights would be interleaved in memory to improve performance on ARM and AVX machines by loading more data in one pass.

Now, there's "online repacking" for weights. Details are in this PR. If you use Q4_0 and your hardware benefits from repacking weights, it will do it automatically on the fly.

As of llama.cpp build b4282, you can't run the Q4_0_X_X files and need to use Q4_0 instead.

Additionally, if you want slightly better quality, you can use IQ4_NL thanks to this PR, which also repacks the weights for ARM (only the 4_4 for now). The loading time may be slower, but it will result in an overall speed increase.

Click to view Q4_0_X_X information (deprecated)

I'm keeping this section to show the potential theoretical uplift in performance from using the Q4_0 with online repacking.

Click to view benchmarks on an AVX2 system (EPYC7702)

model	size	params	backend	threads	test	t/s	% (vs Q4_0)
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp512	204.03 ± 1.03	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp1024	282.92 ± 0.19	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp2048	259.49 ± 0.44	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg128	39.12 ± 0.27	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg256	39.31 ± 0.69	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg512	40.52 ± 0.03	100%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp512	301.02 ± 1.74	147%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp1024	287.23 ± 0.20	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp2048	262.77 ± 1.81	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg128	18.80 ± 0.99	48%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg256	24.46 ± 3.04	83%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg512	36.32 ± 3.59	90%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp512	271.71 ± 3.53	133%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp1024	279.86 ± 45.63	100%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp2048	320.77 ± 5.00	124%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg128	43.51 ± 0.05	111%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg256	43.35 ± 0.09	110%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg512	42.60 ± 0.31	105%

Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation.

Which File Should I Choose?

Click here for details

A great write-up with charts showing various performances is provided by Artefact2 here

First, figure out how big a model you can run. You'll need to know how much RAM and/or VRAM you have.

If you want your model to run as FAST as possible, 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 your system RAM and your GPU's VRAM together, then grab a quant with a file size 1 - 2GB smaller than that total.

Next, 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 (format 'QX_K_X', like Q5_K_M).

If you want more details, check out this extremely useful feature chart: llama.cpp feature matrix

Basically, if you're aiming for below Q4 and running cuBLAS (Nvidia) or rocBLAS (AMD), look towards the I-quants (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 equivalents, so you'll have to decide on the speed vs performance tradeoff.

🔧 Technical Details

The project uses llama.cpp release b5697 for quantization. All quants are made using the imatrix option with a dataset from here.

📄 License

The project is under the iconn license.

Credits

Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset.
Thank you ZeroWw for the inspiration to experiment with embed/output.
Thank you to LM Studio for sponsoring the work.

Want to support the work? Visit the ko-fi page here

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご