Over the past few years, small language models (SLMs) have seen rapid progress in both capability and efficiency. This project, Domyn Edge, aims to push the boundaries of open-source SLMs by developing a model that competes with or outperforms the current state-of-the-art in the 3B–4B parameter range (e.g., LLaMA 3.2-3B, Qwen2.5-3B, Gemma 3, Qwen 3 4B).
A key goal is delivering exceptional performance across multilingual reasoning tasks (English, French, Spanish, German, and Italian), instruction following, agent integration (tool calling), and extended context support (128k+ tokens). The model supports explicit user-controlled reasoning, allowing users to toggle reasoning mode on or off at inference time. When reasoning is enabled, the model engages in deeper logical chains (for math, coding, etc.). When disabled, it focuses on fast, general-purpose conversation. This architecture integrates the latest advancements in attention mechanisms and tokenization to maximise compute efficiency.
The entire project will be released under a permissive open-source license, including model weights, evaluation benchmarks, and datasets. By democratising cutting-edge SLMs, this initiative will foster innovation, inclusion, and broad accessibility across Europe and beyond.
Martin Cimmino, Domyn, Italy