laya-ts browser demo (WebGPU → WASM)

Multilingual checkpoint, fully client-side via onnxruntime-web (WebGPU, WASM fallback). No server inference.

1 Load model idle

Model URL — serves encoder.onnx, head.onnx, tokenizer.json, rl_agent_config.json

First load fetches ~1.3GB, then CacheStorage serves it.

Triage preset loaded. Add your own or remove these.

Type

Instructions (what should the model answer?) Criteria — one per line, label: description

Customer message Ctrl+Enter to run

Load the model to unlock.