Domyn-small

A 10B-parameter open-weight reasoning model that delivers the best accuracy-efficiency balance in the 7-10B class, with dual-mode (thinking on/off) inference, tool calling, and a 32K native context window extensible to 128K.

Domyn Small is a compact, 10-billion-parameter open-weight reasoning LLM designed for high-volume, cost-sensitive use cases. It sets a new standard on the accuracy-efficiency frontier — matching the reasoning quality of peers while spending a fraction of the tokens — bringing serious reasoning capability within reach of enterprises that need to run AI on infrastructure they control. By powering agentic workflows, tool-calling pipelines, and fine-tuned domain applications that fit on a single GPU, Domyn Small enables organizations to deploy AI at scale with the efficiency, sovereignty, and control that regulated industries demand.

Domyn Small is a dual-mode reasoning model with a 32K native context window (extensible to 128K), first-class tool calling, and open weights under the MIT license. With reasoning on it performs extended multi-step reasoning for complex coding, mathematics, and tool-use tasks; with reasoning off it produces faster, more compact answers when depth is unnecessary. Its defining result is token efficiency: against peer models in the 7-10B class it achieves comparable reasoning accuracy while spending roughly 2-4x fewer reasoning tokens, which translates directly into lower inference cost and latency. The model supports 50+ languages — with particular strength in English and the major European languages (Italian, Spanish, French, German) — was trained on European HPC infrastructure (the CINECA Leonardo supercomputer), and is released under the EU AI Act, part of a broader effort to build sovereign-EU AI for regulated industries.