Open-weight non-autoregressive System 1 decision model returning calibrated probabilities in one forward pass across 100+ languages.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Laya (also written Leia) is a family of encoder-based decision models from Convai Innovations. Rather than generating text, it answers typed questions about an input (raw text, email, ticket or JSON) in a single forward pass and outputs only probabilities and numbers. The vendor reports about 33 ms latency on a single GPU and says the output format rules out malformed JSON or free-text hallucination.
It supports three primitives: choice (pick one option from a set of criteria, with a probability distribution and calibrated confidence), score (place the input on an ordinal rubric) and noul (a boolean question returning P(true)). Typical uses are ticket routing, phishing and spam detection, jailbreak/prompt-injection screening, urgency scoring and agent observability. The vendor says it is trained with RLCD and proper scoring rules, and positions it as an open alternative to TypeSafe AI's Jev. Its authors describe it as building on their earlier 2025 arXiv work on RL-based conversion models.
Three checkpoints are bundled in the convaiinnovations/laya Hugging Face repository: an English model on ModernBERT-large (421M parameters, 512 context), a multilingual model on mmBERT-base (322M, 1024 context, up to 8k) covering 100+ languages, and a typed-decisions variant with 1024 context. An SDK can download only the requested subfolder. Fine-tuning scripts are reportedly included. Benchmark figures (including comparisons with Jev) are self-reported by the authors and not independently verified.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source (Apache 2) system-one model similar to Jev that gives confidence scores over multiple-choice options. Its authors say they proposed the concept first, and self-reported benchmarks show it slightly ahead of Jev. It comes in base, multilingual and long-context versions; the main model is 2.37 GB and fine-tuning scripts are included.
Closest neighbours by meaning, not just by category.
Model · Google
Google DeepMind's family of open-weights models built from Gemini technology, from phone-sized to laptop-class variants.
Model · Sarvam AI
Open-weights decision model on Sarvam-105B that returns probabilities for fixed answer options in one forward pass.
Model · Google
Google's open-weights 12B encoder-free multimodal model in the Gemma 4 family, taking text, image and audio input.
Model · Google
Google's open-weights 26B diffusion language model that drafts blocks of text in parallel.
Model · Perplexity
Two open-sourced Perplexity decision models, at 4B and 27B parameters, announced by the company.
Model · Alibaba
Small 0.8B-parameter open-weights language model in Alibaba's Qwen 3.5 family, aimed at lightweight and on-device use.