CygnisAI

Welcome to the official CygnisAI organization on Hugging Face!

About CygnisAI

CygnisAI is an independent AI research organization founded and led by Simonc-44. The organization develops high-performance, open-source language models with a focus on three principles: hardware accessibility, cognitive transparency, and data sovereignty.

Where most frontier AI research concentrates capability behind proprietary infrastructure, CygnisAI inverts this approach — producing models that are fully downloadable, locally executable, and structurally transparent in how they reason. Every model in the Cygnis family exposes its reasoning process through a native Chain-of-Thought architecture, making the decision pathway auditable rather than opaque.

The Cygnis-Alpha 1 Series

The Cygnis-Alpha 1 family is built on top of SmolLM2 1.7B and refined through successive fine-tuning rounds targeting reasoning, instruction following, and identity alignment.

Featured Models

• Cygnis-Alpha-1.7B-v0.1-Instruct — The Flagship. Fine-tuned for chat and complex instructions.
• Cygnis-Alpha-1.7B-v0.1 (Base) — The core weights in Safetensors, ready for fine-tuning or specialized tasks.
• Cygnis-Alpha-1.7B GGUF Series — Optimized for local hardware (llama.cpp, LM Studio).
• Cygnis Alpha 1 on Ollama — Run Cygnis with a single command: ollama run CygnisAI/Cygnis-Alpha-1.7B-v0.1.

Training Assets

• Cygnis-Identity-SFT — The specialized dataset used to align the model's persona and reasoning style.

The Cygnis-Alpha 2 Series

The Cygnis-Alpha 2 family is built on top of Llama 3.1 8B and refined through successive fine-tuning rounds targeting reasoning, instruction following, and identity alignment.

Featured Models

• Cygnis-Alpha-2-8B-v0.2 — Latest The first standalone merged release (no adapter required). Optimized for high-precision instruction following and complex reasoning tasks.
• Cygnis-Alpha-2-8B-v0.1 The original LoRA adapter introducing the unique three-phase reasoning structure. Requires unsloth/meta-llama-3.1-8b-bnb-4bit as a base.

Research & Datasets

• Cygnis-Alpha-2-Instruct-Mix The core instruction dataset. A high-signal mixture of Magpie-Pro and FineWeb-Edu, filtered for maximum cognitive density.

Community and Contribution

CygnisAI is a community-driven research project. Contributions in the following areas are welcome:

• Evaluation runs on additional benchmarks
• Fine-tuning experiments on domain-specific datasets
• Quantization quality reports across hardware configurations
• Documentation translations (French / English)

Feedback, benchmark results, and deployment reports can be submitted through the model discussion pages on this profile.

Open Research Principles

CygnisAI publishes all model weights openly under permissive or clearly stated licenses. The organization does not pursue proprietary lock-in, closed APIs, or usage telemetry. Research outputs are made available without embargo.

For questions regarding commercial licensing, collaboration, or research partnerships, contact Simonc-44 through the Hugging Face profile page.