maple-preview-4bit-onnx-android

Maple Preview 4-bit ONNX for Android

Private OpenMed development export of deepgrove/maple-preview revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07 for ONNX Runtime Mobile. The expert matrices use symmetric 4-bit, 128-value blocks through the fused com.microsoft.QMoE operator. Router computation and the public graph I/O use FP32. The included openmed-maple-onnx-export.json is the integrity receipt for every file, source revision, conversion dependency, and validation gate.

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 KV growth from one to two tokens. The exact fused QMoE form also passes the standalone operator smoke test.

Source-logit parity, useful generation, direct-identifier recall, critical-leakage, peak memory, latency, and physical Android execution remain release gates. This is conversion and CPU-runtime evidence only.

Integration

The cache contract, bundle validator, and Compose demo are documented in docs/maple-on-device.md. Applications must verify the receipt's sizes and SHA-256 digests before moving the selected model into protected app storage. Do not add cloud fallback for PHI workflows.

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.