cstr/OXR-1.0-GGUF

OXR-1.0 GGUF (CrispEmbed)

GGUF conversions of SoMarkAI/OXR-1.0 by SoMark Team, for CrispEmbed's native ggml runtime. All credit for the model belongs to the original authors; these files only change the storage format. They target CrispEmbed's native OXR profile (general.architecture = oxr, short vit. vision tensor names) and are not loadable by llama.cpp. Full-page document parsing additionally uses the D-FINE layout detector published separately as cstr/OXR-1.0-DFINE-layout-GGUF.

License and attribution

• OXR License: Apache-2.0 with additional terms (LICENSE, verbatim; base text in LICENSE-APACHE-2.0). The additional terms take precedence:
  • commercial use without a separate SoMark commercial license is permitted only while you and your affiliates stay below USD 2,000,000 in total funding and in annual revenue;
  • an online service built on OXR or its derivatives must prominently identify SoMark OXR;
  • offering third parties a hosted API that competes with SoMark's document parsing / OCR / layout services is not permitted at any size without SoMark's written license;
  • violating these terms terminates the license automatically.
• These GGUF files are derivative works and are redistributed under the same terms; NOTICE carries the attribution.
• CrispEmbed only loads them after explicit acknowledgement (CRISPEMBED_OXR=1 and CRISPEMBED_ACCEPT_OXR_LICENSE=1); its model registry requires --accept-license oxr to download them, and crispembed-server warns about sections 2 and 3 when an OXR stage is configured.
• UPSTREAM_README.md is the unmodified upstream model card at revision c708407f9d2b4df658f2543d9132b62bb7221df4 (SHA256 a3241e175099ea57a4491be25cce8e965945889e41ed25c3c01911fd17a2236f).

Statement of changes (Apache-2.0 section 4(b))

These files were modified from the original by CrispEmbed: the checkpoint weights of revision c708407f9d2b4df658f2543d9132b62bb7221df4 were format-converted to GGUF (tensor renaming, casting, tokenizer and preprocessing metadata embedded) and, where noted, quantized. No weights were trained, fine-tuned, merged or otherwise altered.

Files

File	Precision	Bytes	SHA256
oxr-1.0-f16.gguf	F16	4,440,553,984	eac8754f4886ba4197416083d49089a3d86c8644f0e5650f44fcf69059d81a57
oxr-1.0-q8_0.gguf	Q8_0-LARGE	2,029,456,896	f0a67a94b9f5416ff35e5c31b7873cea1a15b95743e09a2d032005ead44eca83
oxr-1.0-q4_k-head.gguf	Q4_K-HEAD	1,878,471,168	037a31df623facd71ca96961e48d65d9010bd1b4b16bcd867254d68d57637709

• F16: F16 attention/MLP matrices; embeddings, lm_head, visual vocabulary/head, patch/position tables and norms F32
• Q8_0-LARGE: LLM matrices plus embeddings, lm_head and visual tables Q8_0; SigLIP2 vision tower F16; norms F32. Clears the Q8_0 final-logit cosine bar (> 0.97) on all five cases; the German page differs late (CER 1.1%).
• Q4_K-HEAD: Mixed recipe after fusor-rs/oxr-wasm (MIT): visual head Q4_K, LLM matrices/embeddings/lm_head Q8_0, SigLIP2 vision tower F16. Clears the Q4_K bar (> 0.94) on all five cases.

How the files were produced

1. Downloaded SoMarkAI/OXR-1.0 at the pinned revision c708407f9d2b4df658f2543d9132b62bb7221df4; the converter (models/convert-oxr-to-gguf.py) checks the source files, tensor names and shapes.
2. Quantized variants were produced with CrispEmbed's own crispembed-quantize from the listed source GGUF.
3. Built and uploaded by CrispEmbed commit ed400340b4a071d776f40ff7c0b87ec520bb4d79 (workflow ocr-gguf-publish.yml, protocol docs/oxr-publication-v1.json); manifest.json lists every recipe, input pin, receipt and file hash.

Parity status (CrispEmbed native vs. the original Python model, CPU, greedy)

• The reference arms (f32, f32-reuse) reproduce the original model's generated token IDs exactly on all 5 frozen requests (formula (22), table (261), japanese (38), receipt (202), german (401) tokens), CrispEmbed commit 0e9060fc, reference revision c708407f9d2b.
• F16 is token-ID exact on the same 5 cases (receipt tests/results/oxr/precision.json); Q8_0-LARGE clears the 0.97 final-logit cosine bar on the same 5 cases (worst 0.9977; token IDs not all exact; receipt tests/results/oxr/precision-cosine.json); Q4_K-HEAD clears the 0.94 final-logit cosine bar on the same 5 cases (worst 0.9980; token IDs not all exact; receipt tests/results/oxr/precision-cosine.json).
• Strict per-stage numerical budgets are not met, including F32: one-ulp F32 differences at the decoder's ~2^15 outlier activations exceed the fixed 0.002 absolute budget (worst cosine 0.99999988); decoded IDs and text are exact. Also not verified: GPU execution, mobile/WASM, representative document corpora beyond the five frozen cases.

Usage

OXR support is experimental and requires acknowledging the OXR License:

CRISPEMBED_OXR=1 CRISPEMBED_ACCEPT_OXR_LICENSE=1 crispembed --ocr-pipeline page.png \\
  --ocr-engine oxr --ocr-rec oxr-1.0-f16.gguf