--- title: README emoji: ๐Ÿš€ colorFrom: blue colorTo: red sdk: static pinned: false --- ๐Ÿง  **AxionLab** *Research-driven AI. Fast, smart, and brutally honest.* AxionLab is an independent lab focused on efficient language models, explicit reasoning, and lean architectures that actually run in the real world. *Less* marketing. *More* gradient descent. ๐Ÿ“‰โœจ ๐Ÿš€ **What we do** We create and train AI models with three clear obsessions: Efficiency: small, fast, and affordable models Reasoning: *think better > speak more* Real code: no fragile demos that break outside the laptop If it doesn't run on common hardware, we question it. If it doesn't reason, we redo it. ๐Ÿงชโš™๏ธ ๐Ÿงฉ **Model Lines** Axion Series; our core family of models MiniText โ†’ ultra-fast, minimal cost MiniGPT โ†’ balance between performance and lightness DogeAI-v1.5-Coder โ†’ total focus on small programming DogeAI-v2.0-4B-Reasoning โ†’ *reasoning, logic, and step-by-step thinking* Axion1-350k-A250k โ†’ From scratch 350k parameters MoE SLM, trained on GSM8k MiniAxion1 โ†’ Nano Reasoning Models that surpasses models with the same size! MiniAxion1.5 (next line of thinking models) All focused on open research and incremental learning. ๐Ÿงช **AxionLab Philosophy** Models aren't magic โ€” they're trainable systems Bigger isn't always better Explicit reasoning matters Open-source accelerates science Benchmarks are guides, not gods AI gets better when we question assumptions, including our own. ๐Ÿง ๐Ÿ’ฅ ๐Ÿ“ฆ **What you'll find here** LLM models trained from scratch or fine-tuned Experimental checkpoints Ongoing research Proprietary tokenizers ๐Ÿค**Collab** We collab with: CompactAI(https://huggingface.co/CompactAI) ๐Ÿ’ป **Hardware** RTX 5090 32GB ๐Ÿ’ช๐Ÿ˜(CompactAI) Radeon Vega 7 8GB ๐Ÿฅฒ๐Ÿ‘(With ZLUDA and other techniques)(AxionLab) ๐ŸŒ **Vision** To build truly accessible, auditable, and intelligent AI, not just impressive loss graphs. AxionLab *isn't* about hype. It's about understanding how to think better with machines. ๐Ÿ“ซ **Contact / Collaborations Pull requests, issues, and crazy ideas are welcome.**