Liodon AI

We are an independent U.S.-based research-driven open-source entity dedicated to advancing the frontier of Small Language Models (SLMs) and high-efficiency deep learning infrastructure. Our goal is to democratize highly capable, compact intelligence that can run seamlessly on low-resource settings without sacrificing reasoning ability.

By bridging hardware-aware optimization with innovative architectural paradigms, we design models that maximize performance-per-parameter.

🚀 Core Research Pillars

• Architectural Efficiency: Engineering high-throughput SLMs utilizing advanced Transformer variants, highly tailored Mixture-of-Experts (MoE), and sparse attention mechanisms.
• Extreme Model Compression: Developing state-of-the-art recipes for hardware-aware post-training quantization (FP8, FP4, and sub-4-bit formats), pruning, and structure-preserving knowledge distillation.
• Speech & Modality Fusion: Integrating native Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) layers into ultra-compact foundation models.
• Rigorous Alignment & Optimization: Fine-tuning models using gradient-guided optimization variants and preference alignment to guarantee robustness at scale.

🛠️ Open-Source Ecosystem & Tooling

Beyond our base and instruct models, we maintain specialized repositories and libraries to support efficient deep learning workflows:

• TorchCriterion – A modular library providing advanced loss functions and criteria designed specifically for structured knowledge distillation and model alignment.
• PromptGrad – An optimized framework engineered to propagate gradients through prompt spaces, enabling automated prompt tuning and discrete optimization.
• LoRA-XS – A highly parameter-efficient fine-tuning library designed to optimize low-rank adaptation techniques for ultra-small architectures.

📦 Model Families & Roadmap

Our model releases are systematically categorized across distinct footprints optimized for varying compute profiles:

🤝 Collaboration & Community

We are active contributors to the global AI research community. If you are interested in collaborating on efficient foundation models, data curation for SLMs, or hardware-aware inference infrastructure:

• GitHub: github.com/Liodon-AI
• Discussions: Open an issue or a thread on any of our model cards here on Hugging Face.
• Citations: If you use our models or software frameworks in your academic research, please use the citation metadata available on the respective model repositories.