🧠 ZINI-1-CHAT-STORIES-web

Browser-runnable build of ZINI-1-CHAT-STORIES — the story-only chatbot model. This repo contains the ONNX export (Transformers.js format) used by the demo Space to run the model directly in your browser via WebGPU or WASM — no server, no API quota, no account needed.

• Architecture: Qwen2.5 (0.5B), causal LM
• Base: Qwen/Qwen2.5-0.5B-Instruct (Apache-2.0)
• Source export: onnx-community/Qwen2.5-0.5B-Instruct
• License: Apache-2.0

Included ONNX variants

File	Size	Best for
onnx/model_q4f16.onnx	~483 MB	WebGPU (fast, good quality)
onnx/model_int8.onnx	~512 MB	WASM (any device, slower)
onnx/model_fp16.onnx	~1 GB	WebGPU when max quality is wanted

(Other variants were removed to keep the repo lean; re-add via the source export if needed.)

How the demo uses it

The companion Space (ZINI-1-CHAT-STORIES-demo) loads this repo with Transformers.js:

import { pipeline } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1/+esm";

const device = navigator.gpu ? "webgpu" : "wasm";
const generator = await pipeline("text-generation", "zeene-prod/ZINI-1-CHAT-STORIES-web", {
  device,
  dtype: device === "webgpu" ? "q4f16" : "int8",
  progress_callback: (p) => console.log(p.status, p.progress),
});

const result = await generator(
  "You are ZINI-1, a storyteller. Tell me a bedtime story.",
  { max_new_tokens: 300, temperature: 1.0, top_p: 0.95, do_sample: true }
);

The 0.5B model is small enough to run entirely client-side after a one-time download (~0.5 GB, cached by the browser).

Attribution

Model weights © Alibaba Cloud / Qwen team (Qwen2.5-0.5B-Instruct); ONNX export by the onnx-community project. Redistributed unchanged, Apache-2.0.