08/09/2026
Running AI models efficiently is what decides whether an organization can afford to run them on its own hardware, in its own country, under its own rules.
Our research team took Uno, a technique from the Institute of Foundation Models in Abu Dhabi, and built it into vLLM, the software BroadNet uses to serve models. On the workload
we tested, one GPU produced 74% more output per second than it did before.
That is extra capacity from software alone, on hardware you already own. It matters most for the edge and on-premises deployments our customers ask for.
The full write-up, the research paper and the code are all public:
Uno adds an adapter that proposes several tokens at once. We built it natively into vLLM, the engine we serve on, and measured 74% more output per second than plain vLLM at 32 concurrent streams on an H100. The paper and its evidence are on Zenodo.
