cstr/MiniCPM-OCR-1.0-GGUF

MiniCPM-OCR-1.0 GGUF (CrispEmbed)

GGUF conversions of openbmb/MiniCPM-OCR-1.0 by OpenBMB, for CrispEmbed's native ggml runtime. All credit for the model belongs to the original authors (OpenBMB); these files only change the storage format. They are not the llama.cpp files from openbmb/MiniCPM-OCR-1.0-gguf and are not loadable by llama.cpp.

License and attribution

• The original model is released under the Apache License 2.0, Copyright OpenBMB. These derived files are distributed under the same license.
• LICENSE is the verbatim upstream Apache-2.0 text. The model repository declares license: apache-2.0 but ships no LICENSE file, so it is taken from the upstream code repository OpenBMB/MiniCPM-OCR@8a16a6ad007b (SHA256 186c8635e1f88660ec113f4935867a612e35b3124aca5c712d97e2e33d7eb677). Upstream ships no NOTICE file, so none is included.
• UPSTREAM_README.md is the unmodified upstream model card at revision 6796197bb3a8e224b2f64ae462bfbb3f8a15ac4b (SHA256 a62051d6a7c4cf3bcf6576a6028a12030c9c601f569d9d09ab7bb49502c5294a).

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

These files were modified from the original by CrispEmbed: the BF16 safetensors weights of revision 6796197bb3a8e224b2f64ae462bfbb3f8a15ac4b were format-converted to GGUF (tensor renaming, BF16 to F32/F16 casting, tokenizer and preprocessing metadata embedded). No weights were trained, fine-tuned, merged or otherwise altered.

Files

File	Precision	Bytes	SHA256
minicpm-ocr-1.0-f16.gguf	F16	2,612,763,136	e024db8b1312116f68be6318b42203a1f99db6ddcb60fb1fce672f5e8f1f0362

How the files were produced

1. Downloaded openbmb/MiniCPM-OCR-1.0 at the pinned revision 6796197bb3a8e224b2f64ae462bfbb3f8a15ac4b; every file was checked against the size and SHA256 pins in CrispEmbed's docs/minicpm-ocr-source.json.
2. Converted with models/convert-minicpm-ocr-to-gguf.py (--dtype f32 lossless; --dtype f16; norms, biases and other 1-D tensors stay F32).
3. Quantized variants, if any, were produced from the F32 GGUF with CrispEmbed's own crispembed-quantize.
4. Built and uploaded by CrispEmbed commit 2f2f7ab03e70514c6b2715bcfa8dbf67298e6c7f (workflow minicpm-ocr-publish.yml); manifest.json lists every recipe, input pin, receipt and file hash.

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

• The F32 and F16 conversions reproduce the original model's generated token IDs exactly on all 5 reference cases (line (19), para-en (63), table (56), tall (279), tiny (3) tokens; Transformers 5.21.0, up to 1024 new tokens), CrispEmbed commit 24d516e2.
• Only unquantized precisions are published: no quantized variant has a token-ID-exact precision receipt yet.
• Not verified: GPU backends, multi-page documents, task-level quality benchmarks, prompts other than the reference tasks.

Usage

MiniCPM-OCR support is experimental and default-off in CrispEmbed; it reuses the Qwen3.5 decoder gate:

CRISPEMBED_MINICPM_OCR=1 CRISPEMBED_QWEN35_OCR=1 \
  crispembed -m minicpm-ocr-1.0-f16.gguf --ocr page.png --ocr-prompt "Text Recognition:"