cstr/got-ocr2-crispembed-GGUF

GOT-OCR2 GGUF (CrispEmbed)

GGUF conversions of stepfun-ai/GOT-OCR2_0 for CrispEmbed's native ggml runtime. The original authors get full credit for the model. These files change only the storage format and, where noted, the numeric precision.

Research and evaluation use only. You may not use these files commercially, and you may not use the model or its output to improve any other large language model. See the license section.

Use got-ocr2-f16-r2.gguf. It is the variant CrispEmbed recommends, and the only one whose output exactly matches the original model. got-ocr2-q4_k.gguf is EXPERIMENTAL: its output drifts from the original's on two of the three test cases.

License and attribution

• GOT-OCR2's decoder derives from Qwen/Qwen1.5-0.5B, which is licensed under the Tongyi Qianwen RESEARCH LICENSE AGREEMENT. The upstream apache-2.0 tag cannot relax the license of the base model. These derived files are redistributed under the same research license. Evidence: license audit.
• LICENSE is a verbatim copy of the license text from Qwen/Qwen1.5-0.5B at revision 8f445e3628f3500ee69f24e1303c9f10f5342a39 (7,281 bytes, SHA256 41c2cf8c272f6fb0080a97cd9d9bd7d4604072b80a0b10e7d65ca26ef5000c0c). The GOT-OCR2_0 repository ships no LICENSE file.
• NOTICE is the "Notice" text file that section 2(c) of that license requires. Its first line is the required attribution, verbatim: "Tongyi Qianwen is licensed under the Tongyi Qianwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."
• UPSTREAM_README.md is the unmodified upstream model card at revision 979938bf89ccdc949c0131ddd3841e24578a4742 (3,959 bytes, SHA256 1c8697ddeb841080c25a06b2eadd9dbdf714cdef886fb457a2568b57b2079134).
• CrispEmbed loads these files only when you set CRISPEMBED_ACCEPT_TONGYI_QIANWEN_LICENSE=1.
• An earlier version of this card described the license as apache-2.0, named a Qwen2-0.5B decoder and recommended Q4_K. All three statements were wrong, and this card replaces them.

Statement of changes (license section 2(b))

CrispEmbed modified these files from the original:

• The checkpoint weights of stepfun-ai/GOT-OCR2_0 revision 979938bf89ccdc949c0131ddd3841e24578a4742 were format-converted to GGUF. The conversion renamed tensors, cast them to other types, and embedded the tokenizer and preprocessing metadata.
• The q8_0 and q4_k files were also quantized from the F16 GGUF with crispembed-quantize.

No weights were trained, fine-tuned, merged or otherwise altered.

Files

File	Status	Bytes	SHA256
got-ocr2-f16-r2.gguf	recommended: exact match with the original	1,441,830,112	98b431abc3e6b5715363e7296f7fc665a2eca829127ccb8693511926c4784a7f
got-ocr2-q4_k.gguf	experimental: drifts from the original	445,492,288	ba276f2e246273af7d1bcd58f35ac7bae09c30340b923ed4440effe82d5e7613
got-ocr2-q8_0.gguf	unqualified: not measured against the original	599,632,960	15920499f65b71bcfbf4d44d0e92ffbaf7f7c4523ad13f7b7bef160dc5edf098
got-ocr2-f16.gguf	defective, do not use: kept for provenance	1,441,830,112	aeaf1b44925eeb449584d716f181fc486cdd6becf596ec896276c5d633e996bb
got-ocr2-ref-full.gguf	diagnostic reference dump	170,013,952	27de920f22b7b44c65d4fb32087e51f35b7bc38c9e45d438b2251df05849ffe5
LICENSE	verbatim Tongyi Qianwen RESEARCH LICENSE	7,281	41c2cf8c272f6fb0080a97cd9d9bd7d4604072b80a0b10e7d65ca26ef5000c0c
NOTICE	section 2(c) attribution	391	c3fe8e03056e0610a7f69c38e12066bfe7e4931223d046eaf5596604788ee892
UPSTREAM_README.md	upstream card, unmodified	3,959	1c8697ddeb841080c25a06b2eadd9dbdf714cdef886fb457a2568b57b2079134

The original got-ocr2-f16.gguf contains an 8 MiB block of zero bytes inside l.embed_tokens.weight (token rows 43-2091). As a result it outputs repeated colorcolor... on every backend. got-ocr2-f16-r2.gguf is a fresh conversion. CrispEmbed read back every one of its 471 tensors and found each byte-identical to the original checkpoint.

Quality against the original model

The reference is the original stepfun-ai/GOT-OCR2_0 chat() remote code, run on CPU in F32 with transformers 4.37.2 and greedy decoding. The panel has three cases: fox, receipt_example and simple_form. The figures come from hosted run 41588457181 (receipt).

Variant	Status	Token-ID agreement with the original	First divergent token	Character ratio vs original	Stage cosine vs original (min)	Final-logit cosine
F16-r2	recommended	3/3 cases exact	none	1.0 / 1.0 / 1.0	1.000000 (19 stages)	not measured
Q4_K	experimental	1/3 cases exact (fox)	receipt_example token 83, simple_form token 20	1.0 / 0.8041 / 0.9679	vision 0.997745, decoder 0.995157	not measured
Q8_0	unqualified	not measured	not measured	not measured	not measured	not measured

The Q4_K drift is a real loss of quality, not merely a numerical difference. On receipt_example the text diverges after token 83, and only about 80% of the characters match. Use F16-r2 when results matter. CrispEmbed has not yet run a mixed-precision Q4 search to find a recipe that matches the original. Until one exists, Q4_K stays experimental.

Not verified: GPU backends other than the M1 F16 receipt, multi-page documents, and task-level benchmarks.

Usage

CRISPEMBED_ACCEPT_TONGYI_QIANWEN_LICENSE=1 crispembed --ocr got-ocr2-f16 image.png

Provenance and EU AI Act Art. 53 note

• Upstream model: stepfun-ai/GOT-OCR2_0, published by stepfun-ai. The decoder base is Qwen/Qwen1.5-0.5B.
• What was done here: format conversion and quantization only. There was no training, fine-tuning, merging or distillation.
• Provider status: the upstream authors remain the provider of this model. Training data is documented upstream, if anywhere.

Experimental quantizations from F16-r2 (2026-11-07)

Research/evaluation use only under the Tongyi Qianwen RESEARCH LICENSE (see LICENSE and NOTICE).

file	recipe	size	gold corpus mean CER (55 pages)	token-exact vs F16-r2
got-ocr2-f16-r2.gguf (recommended)	F16	1.44 GB	0.363	-
got-ocr2-q8_0-r2.gguf	Q8_0	600 MB	0.357	17/55 (3/3 on the short panel)
got-ocr2-q8-q4k-mix.gguf	Q8_0, decoder ffn_gate/ffn_up Q4_K	530 MB	0.345	11/55

Both are EXPERIMENTAL: on par with F16 on aggregate gold CER but not token-exact; individual pages can regress. They supersede got-ocr2-q4_k.gguf (kept for provenance). Method and receipts: https://github.com/CrispStrobe/CrispEmbed/blob/main/docs/got-ocr2-q4-research.md