Fine-tune Laya on free Kaggle GPUs
Laya's author recommends fine-tuning for real use cases and shares a 2×T4 Kaggle notebook that any agentic coding tool can adapt.
BUILDS
30 builds made with System One models.
Agents, routers, scorers and games — built on models that decide. Every entry links to the models it uses, so you can pull the same one and start from there.
24 builds
FluidUse: on-device computer use for Mac apps
A Swift package that reads forms in running Mac apps through the Accessibility API and decides what to fill with Laya on the Neural Engine, around 1 ms per decision.
kevala: Laya in any web page with Rust, WebAssembly and WebGPU
A zero-dependency Rust runtime compiled to Wasm with WebGPU kernels; one import loads a pinned int8 model and nothing leaves the browser tab.
laya-go: a Go server, agent CLI and MCP bridge
One process serves the model API and an admin UI; agents ask for decisions over MCP, HTTP or the CLI and act on the top choice.
Laya on a simulated home energy system: fast, not the winner
Four controllers ran the same minutes of a Swedish house. Laya with a safety clip cut the bill, but a tariff-aware energy manager did better; p95 was 179 ms per decision.
laya-mcp (PerryLink): a hardened MCP sidecar for Laya
Adds a token-budget preflight that reports what would be truncated, a persistent calibration store and structured output to a warm MCP and HTTP server.
Laya MLX Swift: Laya inside a native macOS app
A Swift library on Apple MLX with no Python or helper process: download the weights once, then call predict with choice questions.
Laya MPS: decisions on a Mac GPU through PyTorch Metal
Runs the checkpoint on Metal Performance Shaders at about 32 ms median and 2.1 GiB of memory on an M5 Pro.
laya-zig: a dependency-free Zig runtime for CPUs
Tokenizer, encoder and decision head for the 322M multilingual checkpoint in one binary, with no Python, torch or BLAS.
Laya4j: a pure-Java SDK on ONNX Runtime
A Java 17+ port with typed questions, a script-detecting router that picks the right checkpoint, and model downloads from the Hugging Face Hub.
arbiter: a Jev-compatible decision server for Laya or your own model
A self-hosted typed-decision server for NVIDIA GPUs and Apple Silicon that speaks Jev's wire format, with integrations for coding agents.
docker-laya: a complete Laya API service in one container
A FastAPI service that bundles all three Laya checkpoints behind a language router, with API keys, presets, OpenAPI docs and bulk inference.
A fully local Home Assistant conversation agent on Laya
Allen Porter's Assist agent resolves intents and entities in a single Laya pass on CPU, CUDA or Apple Silicon, with nothing sent to the cloud.
laya-candle: Rust inference with custom Metal kernels
Runs the original Laya safetensors through Candle with fused kernels, tiled attention and GEMM dispatch tuned for the M1 Pro.
laya.cpp: a standalone C++ runtime with CUDA and Vulkan
A ggml-based runtime for all three checkpoints with native tokenization, decision heads and JSON output.
laya_ex: Laya in Elixir on Nx and Bumblebee
Fetches the official checkpoint on first use and answers typed questions natively, with EMLX on Apple Silicon.
laya-mcp: laya-mlx for any MCP client
Will Sargent's stdio MCP server lets Claude Code and other clients classify text, score against a rubric or ask yes/no questions locally.
laya-rust: pure-Rust Laya inference
Loads the safetensors directly with candle and runs any checkpoint on CPU, Metal or CUDA — no Python, torch or ONNX export.
laya-web: the whole model in a browser tab
524 MB of 8-bit English weights cached in the tab and run with onnxruntime-web, live at laya-web.pages.dev.
@r4ai/laya-web: client-side Laya with WebGPU
A browser runtime on ONNX Runtime Web that uses WebGPU with a SIMD Wasm fallback and works inside Web Workers.
system-one: typed decisions in TypeScript, as Promises or Effects
One set of question definitions drives a zero-dependency Promise client and an Effect service across Jev, Cloudflare AI Gateway and Laya.
laya-mlx: Laya on Apple Silicon at 60 decisions a second
An MLX port tuned to run in under 1 GB of memory; its demo plays Snake at 60 decisions per second on an M3 Max.
@receptron/laya: Laya from Node.js and TypeScript
An npm package that runs the model with ONNX Runtime, without Python or PyTorch, using the same request shape as a hosted System One API.
Laya: the open-source decision engine
The model itself: typed choice and score questions answered in a single forward pass, Apache 2.0, about 33 ms per question on a T4.
Many of the first builds here were collected by madewithlaya.com. Every entry links to its creator’s own post, repository or site.