Ziroh Labs Launches Kompact AI.
On April 9, 2025, Ziroh Labs unveiled the first version of Kompact AI at IIT Madras. The event showcased Kompact AI's unique capabilities and demonstrated how it redefines AI development on top of CPUs by eliminating the dependency on GPUs.
Industry pioneers graced the launch event.
Kompact AI enables inference and fine-tuning for models with less than 50 B parameters on CPUs. Ziroh Labs has optimised models from DeepSeek, Qwen, and Llama and Phi series to run efficiently on CPUs. These models have been benchmarked with IIT Madras, evaluating both quantitative performance and qualitative accuracy. For the first time, high-performance AI is built and deployed efficiently on CPUs, making AI more accessible, scalable, and cost-effective.
IIT Madras has partnered with Ziroh Labs to establish the Center of Excellence: Impact AI, which is focused on developing practical, efficient AI solutions for real-world challenges. Today, AI development is dominated by "Big AI"—massive, general-purpose models designed for scale.
While powerful, these models often fail to address India's grassroots challenges, as they demand extensive computing power, energy, and infrastructure, making them costly and inaccessible to most communities at the bottom of the pyramid. CoE with IIT-Madras aims to solve the problems at the grassroots level.
Specific use cases—such as aiding farmers, supporting students, or assisting frontline workers in rural areas—are best served by smaller, task-specific models that can run efficiently on smartphones and low-power devices, ensuring accessibility where it matters most. Kompact AI plays a transformative role in this type of use case as it leverages multiple optimisations algorithmically and implementation-wise to execute models on the low end without any support required for external communication, such as the Internet. As a part of CoE, IIT Madras will work on smaller AI models for specific use cases, leveraging Kompact AI to optimise execution, making high-performance AI more efficient, accessible, and relevant to real-world applications.
Watch live stream here
