Laya architecture breaks the monopoly of decision models, an open-source alternative that cuts response time to 33 ms

Enterprise AI production lines typically rely on calling large generative models to perform straightforward classification tasks, such as routing support tickets, sorting phishing emails, or checking code command safety, which imposes unnecessary latency and high inference costs on systems. To address this operational bottleneck, Nandakishore Mukunoth, founder and CEO of "Confl AI Innovations", launched the open-source "Laya" model family under the Apache 2.0 license, designed to serve as an intuitive System 1-style decision engine that is non-self-regressive, achieving ultra-fast response times of up to 32.8 ms on a single GPU and dropping to 7.2 ms per query when processing batched requests.

The Laya architecture replaces long-text generation with direct, mathematically rigorous probabilistic calculations over defined data structures.It bypasses the hallucination problem and the lack of calibration that affect generative models when they predict artificial linguistic confidence scores. The architecture relies on three core components: individual selection from a predefined list of criteria, ranking-based evaluation, and binary logical verification with explicit probability values ranging from zero to one. Because the model’s outputs are entirely numeric values and probability distributions, JSON-file syntax errors or deviations from the programming schema become technically impossible.

The move responds to the launch of the TypeSafe AI lab, co-founded by Diego Almida, one of the contributors to ChatGPT’s development, which offers the closed-source commercial "Jive" model that charges $0.042 per million input tokens and has response times ranging from 150 to 276 ms. Laya outperformed the benchmark tests, delivering execution speeds roughly eight times faster than Jive, and achieving better probabilistic calibration with an error of 0.081 versus about 0.246 for the competing model, while also recording an accuracy of nearly 98 % in phishing detection and exceeding 99 % in spam filtering.

Testing across 51 languages reveals that monolingual decision models fail completely when confronted with non-Latin scripts.The tests showed the English encoder model’s accuracy collapsing to random levels in multiple languages despite reporting spurious confidence scores above 90 %. To bridge this gap, the developer integrated a language router written in Python that explores 22 textual writing systems, including Arabic, Devanagari and Cyrillic, with an operational overhead of less than one millisecond, automatically directing queries to the multilingual "MMBERT" model whose vocabulary spans 256 k tokens and provides effective global coverage for over a hundred languages.

This engineering shift translates directly and practically for software teams and infrastructure managers in Gulf banks, fintech firms, and e-commerce platforms in Saudi Arabia, the United Arab Emirates and Egypt. Replacing costly generative models with lightweight, open-weight sorting and calibration models enables the deployment of content-protection and ticket-verification systems locally within isolated enterprise data centers, eliminating the need to send sensitive data streams to external cloud APIs, thereby lowering inference expenses and meeting the region’s strict data-governance and sovereignty compliance requirements. If you manage enterprise data flows, the logical step is to limit large models to complex cognitive tasks and build fast, first-line defense pipelines that rely on specialized engines to ensure accuracy and cut response times to fractions of a second.