Decision-model landscape
System One Models & Typed Decision Engines
Laya is part of a fast-growing model category focused on typed decisions rather than long-form generation. This page tracks the different approaches without collapsing them into one ranking.
| Model / approach | Status | Access | Deployment | Approach | Languages | Source |
|---|---|---|---|---|---|---|
| Laya Open-source System 1 decision model family with Router, local inference, fine-tuning and multiple runtimes. | Open weights | Apache 2.0 project | Local / self-hosted | Non-autoregressive typed decision model with choice, score and noul heads | English + multilingual checkpoint | NandhaKishorM/laya |
| TypeSafe Jev Hosted System One decision model that established much of the current typed-decision API vocabulary. | Hosted model | Managed API | Provider-hosted | Typed choice, score and noul decisions through a hosted model API | Provider-dependent | TypeSafe documentation |
| AnyJev Nokia Applied Research project that turns compatible LLMs into Jev-style typed decision models without task-specific training. | Open library | Apache 2.0 | Bring your own LLM backend | Training-free typed decision layer over existing LLM logits with debiasing and calibration levels | Depends on underlying LLM | nokia-applied-research/AnyJev |
| Bespoke Nimble Bespoke Labs publishes a model, training methodology, evaluation tools, and recipe for open typed decisions. | Open model + recipe | Open repository | Apple Silicon / NVIDIA | Jev-inspired local typed decisions plus data curation and training recipe | Model-dependent | bespokelabsai/nimble |
| OpenThai-SystemOne Thai and English System One model with a Jev-compatible request/response contract and an open training pipeline. | Open weights | Apache 2.0 | CUDA / MPS / CPU | Qwen3.5-0.8B text tower with a 256-way slot head for choice, score and noul | Thai + English | iapp-technology/openthai-systemone |
| Decider A growing family of open System One-style models with training data, evaluation, serving, games, and vision experiments. | Open model family | Apache 2.0 | Local / HTTP serving | Qwen3.5-based one-pass typed decisions with calibrated probabilities and serving engines | Model-dependent | Mapika/decider |
What makes a System One decision model different?
The common idea is to answer structured questions about a state without generating a long free-form response token by token. Applications ask for a label, ordered score, probability, or similarly constrained output and receive machine-readable results.
The projects above are not interchangeable. Some are specialized trained models, some are hosted services, and some are libraries that read or calibrate logits from an existing LLM. Compare deployment model, calibration, task distribution, language coverage, option count, latency, and reproducibility before choosing an approach.
laya-ai.com remains focused on Laya. This category page exists to explain the ecosystem around it and to track adjacent techniques as the decision-model space evolves.