Cygnis Alpha 1.7B - GGUF Collection (i1 & imat)

Welcome to the centralized repository for quantized versions of Cygnis Alpha 1.7B. This repository hosts various optimized variants (GGUF) for local inference, ranging from ultra-compressed formats to high-fidelity versions.

Base Model: cygnisai/Cygnis-Alpha-1.7B-v0.1
Architecture: Based on SmolLM2-1.7B (Llama-style)

Quantization Table

All files below include an Importance Matrix (imatrix) to significantly improve coherence and reasoning at lower bitrates.

File (.gguf)	Method	Size	Best For
`cygnis-alpha-1.7b-v0.1-q5_k_m-imat.gguf`	Q5_K_M	1.23 GB	Recommended - High precision
`cygnis-alpha-1.7b-v0.1-q5_k_s-imat.gguf`	Q5_K_S	1.18 GB	Great balance of speed/quality
`cygnis-alpha-1.7b-v0.1-q4_k_m-imat.gguf`	Q4_K_M	1.06 GB	Standard usage
`cygnis-alpha-1.7b-v0.1-q4_k_s-imat.gguf`	Q4_K_S	999 MB	Fits easily in 1GB VRAM/RAM
`cygnis-alpha-1.7b-v0.1-iq4_nl-imat.gguf`	IQ4_NL	991 MB	Optimized 4-bit coherence
`cygnis-alpha-1.7b-v0.1-iq4_xs-imat.gguf`	IQ4_XS	940 MB	Lean 4-bit performance
`cygnis-alpha-1.7b-v0.1-iq3_m-imat.gguf`	IQ3_M	810 MB	Significant compression
`cygnis-alpha-1.7b-v0.1-iq3_xxs-imat.gguf`	IQ3_XXS	680 MB	Extreme light-weight usage

Quick Start

With Ollama

You can use these models by creating a Modelfile:

FROM ./cygnis-alpha-1.7b-v0.1-q5_k_m-imat.gguf
PARAMETER temperature 0.7
SYSTEM "You are Cygnis Alpha, a sovereign AI designed by Simonc-44. You are polite, fast, and concise."

With llama.cpp (CLI)

./llama-cli --hf-repo cygnisai/Cygnis-Alpha-1.7B-v0.1-i1-GGUF \
            --hf-file cygnis-alpha-1.7b-v0.1-q5_k_m-imat.gguf \
            -p "Hello Cygnis, introduce yourself."

About Cygnis Alpha

Cygnis Alpha is a Sovereign AI developed by Simonc-44. It has been fine-tuned via SFT (Supervised Fine-Tuning) for:

Identity Alignment: It is fully aware of its origin and creator.
Conciseness: Designed for fast, direct, and helpful responses.
Efficiency: Operates on almost any hardware due to its compact 1.7B parameter count.

License

This project is distributed under the Apache-2.0 license. Quantizations were performed using GGUF-my-repo.

Citation

@misc{allal2025smollm2smolgoesbig,
      title={SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model}, 
      author={Loubna Ben Allal and others},
      year={2025},
      eprint={2502.02737},
      archivePrefix={arXiv},
}

Contact & Creator: Simonc-44