InfraOCR models (SoMark OXR-1.0 for Snapdragon NPUs)

QNN HTP context binaries of SoMark OXR-1.0 for the InfraOCR Android app, built with QAIRT 2.51.40.260930. They are derived from SoMarkAI/OXR-1.0 and are distributed under the same license, Apache-2.0 with SoMark's additional terms (see LICENSE and NOTICE).

Folder	Contents
common/	embed.f16 (Qwen3 embedding table), prefix_kv.f32 (K/V of the fixed chat prefix), vte_indicators.f32, pos_embed.f32 (32x32 vision position embedding), vocab.bin (id to bytes), prompts.json (token ids of the fixed prompts)
sm8850/	Snapdragon 8 Elite Gen 5
sm8950/	Snapdragon 8 Elite Gen 6
sm8975/	Snapdragon 8 Elite Gen 6 Pro
app/	InfraOCR APKs and latest.json for in-app updates

Each SoC folder has:

File	Model	Precision
oxr_layout.bin	D-FINE B2 layout detector, 960x960	FP16
oxr_vis_1024.bin, oxr_vis_3600.bin	SigLIP2-NaViT encoder + visual head, padded to 1024 / 3600 patches	FP16
oxr_llm.bin	Qwen3-0.6B decoder, AR-16 prefill and AR-1 decode graphs	W8, A16, KV8
oxr_llm_head.bin	final norm + tied lm_head	W8

manifest.json lists every file with its size and sha256.

Measured against the bf16 reference on 495 blocks from 40 OmniDocBench pages: vision 99.5% exact match at the block level, LLM 97–98% exact match, about 145 ms per vision block and 45–60 tokens per second on SM8975.

License notes

Commercial use without a separate license from SoMark is allowed only while you and your affiliates have raised at most 2 million USD and had at most 2 million USD of revenue in the last fiscal year. OXR may not be used to offer a hosted document parsing, OCR, layout or extraction API. Read LICENSE for the full terms.