Presented is a non-autoregressive model openJev-verdict-2.0

Developers introduced openJev-verdict-2.0 — a non-autoregressive decision-making model designed to work with corporate business processes. The release is focused on high-speed data processing while maintaining accuracy comparable to larger analogs.

The model's architecture includes 149.6 million parameters. During testing, the system demonstrated an accuracy of 77.10% with an expected calibration error (ECE) rate of 1.44%. The inference time is between 20 and 25 ms per request, allowing the solution to be used in high-load real-time systems.

According to the results of internal tests, the model outperforms solutions like Laya in terms of efficiency when performing standard corporate tasks. Developers note that the use of a non-autoregressive approach significantly reduced the delays characteristic of traditional generative models.

The project's code is published openly on GitHub. The model is already available for integration into existing workflows that require minimal response time for automated decision-making.