Laya

Tony Dinh ran laya-mlx on a MacBook against Jev in real-time Tetris. Laya decided far faster but stacked itself to the ceiling; Jev won every round.

Adds image understanding to Laya using a small SmolVLM-256M backbone, so you can ask calibrated yes/no, choice and score questions about a picture plus text.

Describe a past conversation in plain language and Chat Seek finds it across local agent histories, reranks with the receptron Laya runtime, and offers a Resume action.

Simplifying AI ran the identical Snake and typed decisions for 30 s on each model. Laya (421M, local) scored 46 at 86.5 decisions/s; Jev 1.13.0 over the API scored 1 at 3.2/s.

A maze demo on a base M5 Mac where local Laya answered in about 19 ms median against roughly 300 ms for the hosted Jev API. Shows the latency gap that on-device decision models open up.

Typed-decision server on NVIDIA GPUs or Apple Silicon that speaks the Jev wire format, with coding-agent integrations. Built after a cat woke the author at five on a Sunday.

Nerve gives Hermes agents typed System One decisions, ranking, verification and an opt-in tool gate, using Jev or the open Laya model to oversee spending and risk.

Tom Siwik’s point: the model matters less than how you define the decision space and batch. Jev wins networked; Laya-mlx and his own JAX build tie locally.

Runs the typed-decisions checkpoint on Metal Performance Shaders: ~32 ms median and ~2.1 GiB RAM on M5 Pro, plus a slower ~0.74 GiB low-memory mode. Demo and benchmark included.

Local decider for the browser-agent loop: hand it a numbered table of page controls and it picks the operation and element. Playwright/CDP driver, speaks /v1/systemone.

An open-source router that asks Jev or a self-hosted Laya how hard each prompt is, then picks the model and reasoning effort so simple turns stay cheap.

Convert → quantize → deploy path that turns Laya (and Kev, NanoJev, PlayJev) into an INT8 ONNX service on the official wire protocol. Runtime is onnxruntime, tokenizers, numpy.

A single Rust binary that lets you ask questions about log lines, diffs or JSON and explore the answers live, running Laya on CPU with no server.

Shows Laya's decisions running in WebAssembly to route steps of a workflow, so branching logic works without a hosted model.

Elixir library that downloads the official checkpoint on first load and answers typed choice and noul questions; pick EMLX on Apple Silicon or EXLA for CPU/NVIDIA.

No Python, no torch, no ONNX export. Loads the ungated safetensors directly and runs any of the three checkpoints on CPU, Metal or CUDA.