Hosted Laya API
Laya is open weights, so you can always run it yourself. A hosted API makes sense when you do not want to run and patch a model server, when your traffic is small or spiky, or when you want to compare Laya with other decision models before committing. This page lists the options and what to check.
419 projects
laya_dart
manukj
Runs the Laya decision model locally in Flutter and Dart applications
laya-serve
c4bbage
Serves Laya inference with Go, dynamic batching, and TensorRT or CUDA execution
laya-node
roryyu
Provides a Node.js SDK for local Laya typed decisions using Transformers.js and ONNX
laya-infra
ulevitsky
Deploys Laya and its playground with Docker Compose or Kubernetes
dsh-laya
RestCode72
Adds Laya typed-decision tools and a service to DeepSeek Harness
laya-finetuning
tehtommeh
Runs Laya locally with an API and UI, serving multiple checkpoints through the official SDK
laya-api
hugomes14
Serves a multilingual ticket classification model through an HTTP API
laya-http
shaul1991
Serves local Laya predictions through an authenticated self-hosted HTTP API
Laya-Decision
GTC6244
Ports Laya inference to Rust with native Candle support, a router, HTTP server, and CLI
laya-js
johnhenry
Runs Laya typed decisions in JavaScript across MLX, WebGPU and CPU backends
laya-server
cmoro-deusto
Serves Laya System 1 decisions through a Jev-compatible REST API
laya-cli
CHENG-LIANG1
Wraps Laya typed decisions in a stable JSON command-line interface
laya-ts
mikeboe
Serves the Laya model over an API and benchmarks it against the hosted Jev API
laya-service
idCntrue
Exposes the Laya non-autoregressive decision engine through a production HTTP service
laya-mlx-mcp
Blackphi6
Exposes Laya-MLX typed decisions through an MCP server for Apple Silicon
codex_laya_dev
sirogarasu
Provides a Codex devcontainer and GPU API sidecar for running the Laya model
laya-decision-engine
vampirethoran
Serves local Laya typed decisions over HTTP with multilingual routing and inference metrics
laya-kev-docker
AndiMajore
Deploys Laya and Kev decision servers in configurable CUDA-enabled Docker containers
laya-system-one
italoalmeida0
Provides a local JavaScript Laya decision engine with WebGPU and WASM acceleration
laya-r-mlx
tjameswilliams
Provides a native Rust inference engine for Laya with MLX and candle backends
jev-systemone-local
katya4oyu
Plans a local Jev-compatible System One server with MLX, Core ML, and ONNX Laya backends
jev-to-laya-proxy
diwakersurya
A Bun proxy exposes the TypeSafe Jev API and forwards requests to a local Laya server
jev-kev-laya-selfhost
suarify
Provides a Dockerized, self-hosted HTTP API for Laya typed decisions with agent-skill documentation
ruby_llm-providers-laya
codenamev
A RubyLLM provider uses local ONNX Laya checkpoints to answer judge questions
Disclosure: layahost is run by the same team as laya.tools. It is listed first for that reason, not as a ranking.
Self-host or hosted?
Self-hosting is the better fit when text must not leave your machines, when volume is high and steady so that your own hardware is cheapest per call, or when you need on-device latency. The API servers page covers laya-serve, ollaya, Docker and native servers, and the run-locally guide covers Python, Mac, Node.js and CPU setups.
Check the hardware first. laya-cpu-benchmark concluded that a realistic CPU call is 0.3 to 2 seconds, and the Railway template reports about 0.2 to 0.9 s per call on CPU, with around 3.4 GB of memory for both checkpoints. Without a GPU or an Apple Silicon Mac, a self-hosted server can be slower than a hosted one.
A hosted API is the better fit when you do not want to operate a model server, when traffic is low or bursty, or when you want features around the model, such as templates, test sets or other models on the same endpoint.
Hosted options
Details are from each service's site, checked on 2026-10-14. Neither is affiliated with Convai Innovations or TypeSafe.
layahost
- Models: Laya (
laya-english,laya-multilingual) on its own servers in Helsinki. Cloudflare's open Clef and Clef-flash on the same API, answered on Cloudflare's Workers AI GPUs. - API: the Jev-compatible
POST /v1/systemone, and a simpler/v1/decidewith ready templates: moderation, spam, support_intent, routing, sentiment, priority, prompt_injection, pii and language. - Tools: test sets in the console to compare models on your own labelled data, decision flows built in a visual editor, an MCP server, an n8n node (
n8n-nodes-layahost) and a drop-in for Google's Perspective API. - Pricing: Laya per decision (one answered question), the same whatever the text length. Clef per input token. New accounts get $1 of credit after email verification, with no card.
- layahost.com · migrating from Jev
Laya Studio
- Models: Laya's English and multilingual checkpoints, routed automatically.
- Location: GPUs in Switzerland and an account database in Zurich. A Swiss-only mode keeps requests from failing over abroad. The site says request content is not stored and request metadata is kept for 30 days.
- API:
POST /v1/systemoneatapi.laya.studio. - Pricing: per input token, like Jev. New workspaces start with 6k free tokens.
- laya.studio
Railway template
The Railway template is a middle way: a one-click deploy of a Jev-compatible Laya server into your own Railway account. You pay Railway for the machine and your text goes only to your own service, but running and updating it is still your job. Its page says a Hobby or Pro plan is needed for the memory.
What to check
- Data location. Where the model runs, where accounts are stored, whether request text is kept, and whether a failover can move a request to another country. On layahost, Laya and Clef run in different places.
- Pricing model. Per-decision pricing does not depend on text length, so a long email costs the same as a short one, but three questions in one request count as three decisions. Per-token pricing, which Jev also uses, grows with the length of the text. Price both with your real message length and questions per request.
- Jev compatibility. Both hosted services accept
/v1/systemoneand say an existing Jev client works after a base-URL and key change (layahost documents this for the official TypeSafe SDKs). The answers still come from Laya, so a confidence threshold tuned on Jev does not transfer. See the differences on the API servers page. - Latency. The network adds to the model's own time. Laya Studio reports about 120 ms end to end from Europe for a warm request, and layahost lists about 0.1 s for Laya and 0.4 to 1 s for the Clef models. Measure from your own region.
- Accuracy. Hosting does not change the model. Independent benchmarks put base Laya behind Jev and the larger open alternatives on hard tasks (see the benchmark section there, and Laya benchmarks for the model card's own numbers). Run a few hundred of your own labelled examples before you move production traffic.
More ways to use Laya