Qwen3.5-0.8B ONNX (q4f16) for edgextract WebGPU

Browser-ready ONNX slice used by the edgextract demo’s WebGPU in this tab backend.

What this is

• Architecture: Qwen3.5-0.8B (model_type: qwen3_5)
• Quant: q4f16 (embed_tokens, decoder_model_merged, vision_encoder)
• Source graph: onnx-community/Qwen3.5-0.8B-ONNX (Apache-2.0)
• Base weights: Qwen/Qwen3.5-0.8B

This is not the Together Tev1 fine-tune. Tev1 ONNX export is blocked until Optimum supports qwen3_5 from that checkpoint. Use this graph for in-tab letter-logit scoring; use Ollama tev1 when you need Tev1-parity scores.

Files

Only the Transformers.js dtype: q4f16 siblings are published (~650 MB):

• config.json, tokenizer, processor metadata
• onnx/*_q4f16.onnx (+ .onnx_data)

Load in Transformers.js

import { AutoTokenizer, Qwen3_5ForConditionalGeneration } from "@huggingface/transformers";

const model_id = "raphaelmansuy/qwen3.5-0.8b-onnx-webgpu";
const tokenizer = await AutoTokenizer.from_pretrained(model_id);
const model = await Qwen3_5ForConditionalGeneration.from_pretrained(model_id, {
  device: "webgpu",
  dtype: {
    embed_tokens: "q4f16",
    decoder_model_merged: "q4f16",
    vision_encoder: "q4f16",
  },
});

In the edgextract demo, keep the WebGPU model id as raphaelmansuy/qwen3.5-0.8b-onnx-webgpu and press Load Tev1.

License

Apache-2.0 (inherited from Qwen3.5-0.8B / the onnx-community export).