ConvAI Innovations released the open model Laya

ConvAI Innovations company presented Laya — a family of non-autoregressive models for decision-making, positioned as an open alternative to the Jev architecture. The novelty is focused on tasks of the "System 1" type and provides a performance boost of 6–8 times compared to its predecessor.

The models are distributed under the Apache 2.0 license, which eliminates the costs of using the API and allows them to be deployed locally. The release includes three specialized checkpoints, optimized for performing specific logical operations: choice, score, and boolean primitives.

The technical implementation of Laya excludes the autoregressive approach, which allows the model to process requests much faster than standard LLMs while maintaining accuracy in decision-making tasks. The absence of the need for cloud computing makes the tool suitable for integration into systems with strict latency requirements.

This release expands the ecosystem of tools for high-speed data processing, started with the development of the Jev architecture. Full documentation and model weights are available on the project's official website, and discussion of technical aspects of the implementation is ongoing in the Hacker News community.