The solution contains 219 million parameters, which is approximately five times fewer than the original model with around 1 billion parameters. It operates approximately 10 times faster while maintaining accuracy with a deviation of less than 1%.
The AI laboratory of Kryptonite developed the Kryptonite model using knowledge distillation. In this process, a larger trained neural network acts as a teacher to a more compact model, transferring not only correct answers but also the confidence level for each option.
The model analyzes acoustic and phonetic features of speech directly from the audio signal. It provides identification results as standardized language codes such as rus and eng, which simplifies integration into speech recognition data processing pipelines.
The development is intended for telecom operators, technical support services, and multinational companies processing millions of audio recordings. This approach helps reduce the load on data centers and avoids the proportional increase in the number of servers.
In 2026, the average monthly cost of renting dedicated GPU infrastructure for AI in Russia grew by 64% compared to 2025. Server rack placement costs in Moscow increased by 15.7% to 170.8 ₽ thousand, while in Saint Petersburg they rose by 31.3% to 138 ₽ thousand.
