@receptron/laya
Node 20+ ONNX Runtime client for Convai Laya: typed systemOne, HF cache about 1.7 GB fp32, parity claims vs Python RLAgent.
Overview
receptron/laya (github.com/receptron/laya, npm @receptron/laya, MIT) implements the Convai Innovations Laya decision model for Node using onnxruntime-node. Laya.load() fetches about 1.7 GB fp32 weights into ~/.cache/receptron-laya unless LAYA_CACHE overrides. systemOne batches all questions in one run with Choice, Score, and Noul outputs matching Python RLAgent.system_one and TypeSafe system_one shape to four decimal places per README. Model weights remain Apache-2.0 upstream at convaiinnovations/laya; receptron publishes ONNX bundles such as receptron/laya-onnx. This listing is the JavaScript runtime only. Do not treat it as a second Laya weights product page.
Problem: Node services and edge workers cannot import Python RLAgent, yet teams still want Convai Laya probabilities without standing up a separate inference container.
Built for: TypeScript backends on Node 20+ that need Laya systemOne parity, Hugging Face download caching, and ONNX Runtime inference without PyTorch.
First indexed on Jev Directory: 2026-09-27
Creator and team
- Name
- receptron
- Handle
- @receptron
How Jev is used
- Role in the product flow
- Local ONNX System One client for parallel questions over structured state in Node
- Primitives
- ChoiceScoreNoul
- State in
- Plain objects such as ticket subject and body fields or arbitrary JSON state passed to systemOne.
- Decision out
- Typed answers map with choice probabilities, score expectations, or noul P(true) per question key.
- npm install @receptron/laya on Node 20+
- await Laya.load() to download or load local ONNX bundle
- Call systemOne with state and questions record
- Read per-question typed answer objects without casting
- await laya.close() when shutting down the process
ollaya-dev-ollaya is the Rust daemon story: pull many open decision checkpoints and serve POST /v1/systemone on port 11435. receptron-laya is the embeddable npm library when a single Node service needs Laya only. jaredpalmer-kev and wfzyx-von remain separate open model families with their own repos. Point model authorship to Convai Innovations Laya; point package maintenance to receptron. For browser or agent automation, cross-read browser-use-jev-ultrafast instead of forcing Laya into the hot path.
Sourced performance claims
- README documents about 140 ms for three questions on warm Apple-silicon CPU and about 1.7 GB fp32 download size.Source: github.com/receptron/laya README
- Public GitHub repo receptron/laya had 500 stars and 44 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Typed systemOne API matching Python reference
- HF download cache with progress callbacks
- Optional local modelDir from export/export_onnx.py bundle
- Batched questions in one ONNX forward pass
- executionProviders and sessionOptions tuning
Stack
- Node.js 20+
- TypeScript
- ONNX Runtime
- Hugging Face hub downloads
Pricing: MIT package; upstream Laya weights follow Convai license; you pay for compute and bandwidth.
Links
FAQ
- Does this re-list Laya model weights?
- No. We document the receptron Node runtime. Weights and training story stay on Convai Innovations Hugging Face repos.
- How does this differ from ollaya?
- ollaya is a multi-model pull and serve CLI. @receptron/laya is a library for embedding Laya in Node apps.
- How much disk and RAM?
- README cites about 1.7 GB fp32 cache plus a few hundred MB per batch on top of roughly 2 GB loaded model budget.
- Can I pin ONNX revisions?
- Laya.load accepts repo, subfolder, and revision parameters documented in README Options.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.