We study speech, language, agents, and efficient systems from Nepal. The constraints here are real, and solving under them produces methods that travel.
Models built around how Nepali is actually spoken and written, including dialects, code-switching with English, and Devanagari text. Our first result is Kriti V1, an open speech recognition model that holds the lowest raw word error rate across nineteen evaluated systems.
Read about Kriti V1Systems that plan, use tools, recover from failure, and finish long tasks reliably. We care most about how agents behave outside English and outside the environments they were trained in.
Learning systems designed around Nepali classrooms, teachers, and students, not translated from products built for somewhere else.
Open datasets, honest benchmarks, and the infrastructure to reproduce results. When existing benchmarks miss the real task, we build the measurement first.
When existing metrics miss the real task, measurement becomes part of the research, not an afterthought.
A useful result includes where the system breaks, what it costs, and which assumptions are doing the work.
A strong claim comes with artifacts: weights, code, evaluations, and exact numbers anyone can check.
We treat compute as a hard constraint. Efficiency increases who can reproduce, challenge, and extend the work.
We collaborate with researchers, universities, and institutions who care about multilingual intelligence and rigorous real-world evaluation.
research@naamchelabs.com