Laya is an open-source decision model with 421M parameters, first released in March 2025 with an arXiv paper, Hugging Face weights and an open dataset, built around non-autoregressive decisions and sub-35ms latency. Recent coverage focuses on how it performs in practice: it was the fastest of four models benchmarked on an RTX 4090 at 3.9 ms per word with 97.4% accuracy, gained OpenVINO support for roughly 40 ms per question on CPU, and was tested in a Doom control benchmark and served by local decision servers such as Ollaya.
A developer open-sourced Vega, a physics-based typed decision model with only 800M parameters (a 4B version also exists), 73k context and image support, surpassing many Jev benchmarks in single-shot tests. It uses an engine+adapter test-time training architecture where outcomes are modeled as valleys and a friction-affected ball settles on the best result.
A user benchmarked four new open decision models, Laya, d1 3B, Clef-Flash 9B and Lev 4B, on a single RTX 4090, tasking them to flag centipede names word by word across 9,534 words of Wikipedia text. Laya was fastest overall at 3.9 ms per word, about 7,980 words in 32 seconds and 97.4% accuracy; Lev was slowest but caught the most names.
A Reddit user benchmarked six decision models by having them play Pac-Man in real time and published an open-source leaderboard. jev 1.13 led with an average score of 2,750, followed by GPT-6 Luna and Clef Flash. Users can also play as Pac-Man against model-controlled ghosts.
Tests show GPT-6 Luna, behind OpenAI's new Decisions API, is poorly calibrated: when it claimed 99% confidence across 3,600 reasoning problems, it was right only 68% of the time, highlighting AI confidence-calibration problems.
A study crafts input prompts so the first output token answers the question, enabling standard LLMs such as GLM-5.3-Flash to make decisions in a single forward pass and reproduce Jev-like properties on vLLM. Benchmarks show accuracy and speed on par with Jev and substantially better than Laya, though cost per decision is several times higher than Jev; the setup also supports vision inputs.
Ollaya is an open-source decision-model server for local hardware that answers in a single forward pass instead of token-by-token generation, handling a five-question request in about 10 ms. It exposes TypeSafe-compatible endpoints and works with the official Python SDK 0.7.1, shipping open models such as Laya, decider and nli.
Developer Rupesh added OpenVINO support to Laya, achieving 40 ms per question on CPU, 3.4x faster than PyTorch. The project is open-sourced on GitHub with a Flappy Bird demo.
A user converted the 421M-parameter Laya model to OpenVINO and quantized it to INT8, running Flappy Bird locally on a 12th-gen Intel Core i7 CPU, showing small-model inference on consumer CPUs.
jevals is an open-source tool that replaces LLM judges with typed Jev decision models for agent evals and guardrails. All evals for a trace are sent as one HTTP request costing a few thousandths of a cent and returning in a few hundred milliseconds, suitable for every trace or inside the agent loop. It supports Jev, Kev, Laya, and regular chat LLMs.
An experiment let Jev, Laya, finetuned ModernCE and finetuned Qwen3.5-4B play ViZDoom autonomously, comparing kills, survival time and latency. Finetuned Qwen3.5-4B averaged 3.63 kills, the best, while Jev survived 5.63s on average with p95 latency near 200ms. No model received Doom-specific training.
A developer says he released an RL-based non-autoregressive decision model, Laya, back in March 2025 via an arXiv paper, Hugging Face weights, and an open dataset—sub-35ms inference across 100+ languages. He criticizes frontier lab TypeSafe AI, founded by a former OpenAI member, for launching a similar product, Jev, in September 2026 as a brand-new breakthrough without papers, open weights, or training data.
A Reddit poster released Laya, a 421M‑parameter open‑source model trained on a single RTX 6000 Pro, claiming it surpasses Jev benchmarks; the author published a Hugging Face demo and GitHub repo for testing and reproduction.
That is everything
The A.I. Daily
Open-source Vega: 800M Physics-based Decision Model with 73k Context and Image Support
