---
base_model: deepgrove/maple-preview
base_model_relation: quantized
language:
  - en
library_name: onnxruntime
license: mit
pipeline_tag: text-generation
tags:
  - onnx
  - onnxruntime-web
  - webgpu
  - browser
  - maple
  - mixture-of-experts
  - quantized
  - openmed
---

# Maple Preview 4-bit ONNX for browser WebGPU

Private OpenMed development export of
[`deepgrove/maple-preview`](https://huggingface.co/deepgrove/maple-preview)
revision `ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07` for a compatible ONNX
Runtime Web/WebGPU integration. The expert matrices use symmetric 4-bit,
128-value blocks through the fused `com.microsoft.QMoE` operator; cache and
logit I/O use FP16. The included `openmed-maple-onnx-export.json` records file
hashes, source provenance, the exact toolchain, and validation gates.

## Validation status

The complete 24-layer graph passed ONNX checker and OpenMed's graph contract.
On ONNX Runtime 1.25.1 CPU it completed a real one-token prefill and cached
decode with finite logits and correct KV growth. Physical browser WebGPU
execution, source-logit parity, useful generation, direct-identifier recall,
critical-leakage, peak memory, and latency remain release gates.

This custom Maple graph is not a stock Transformers.js model. The browser demo
requires an audited, same-origin adapter with compatible QMoE support. See
[`docs/maple-on-device.md`](https://github.com/maziyarpanahi/openmed/blob/master/docs/maple-on-device.md)
and the standalone demo in `docs/demo/web/`. Notes and generated outputs must
not be persisted in browser storage, and PHI workflows must not use cloud
fallback.

Maple Preview is a research model, not a medical device. It must not
automatically trigger clinical or disclosure decisions. The upstream model is
licensed under MIT; see its repository for the license and model card.