cstr/MinerU2.5-Pro-2604-1.2B-GGUF

MinerU2.5-Pro-2604-1.2B GGUF (CrispEmbed)

GGUF conversions of opendatalab/MinerU2.5-Pro-2604-1.2B by the MinerU Team (OpenDataLab), 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 MinerU profile of its Qwen2-VL graph and are not llama.cpp mmproj pairs.

License and attribution

• The original model is released under the Apache License 2.0 (Copyright 2024-2026 MinerU Team (OpenDataLab)). These derived files are distributed under the same license.
• LICENSE: the model repository declares license: apache-2.0 but ships no LICENSE file, so this is the publisher's own LICENSE.md from opendatalab/mineru-vl-utils@c9797cde8f54 (SHA256 0b0c9cbb9e9f032856deef12e761c0661f14fdbaf88753bb1e1c68d1f7e5a3da). Upstream ships no NOTICE file, so none is included.
• UPSTREAM_README.md is the unmodified upstream model card at revision d3f5e08d073c21466bbabe21c71bb1e9c2e595da (SHA256 c7859313699e858a61074e7d2ab171bea25d37504f3d530b9e0d4232c91706a5).

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

These files were modified from the original by CrispEmbed: the safetensors weights of revision d3f5e08d073c21466bbabe21c71bb1e9c2e595da 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
mineru25-pro-2604-f16.gguf	F16	2,591,511,520	7d464d375f7c20dd3e2bae7e20fed1f80485ae9ca69aaad201897652c74d746b
mineru25-pro-2604-q8_0.gguf	Q8_0	1,920,768,000	0c3f9eaf0e916fe85546fac47267677502241a2b0c3ea76b5e949cfd193601f4

• F16: F16 weights, F32 norms/biases
• Q8_0: decoder and embeddings Q8_0; vision tower F16 (crispembed-quantize MinerU policy, as for InternVL2: Q8_0 vision spiked to a 0.49 worst-row cosine on the CPU stage diff)

How the files were produced

1. Downloaded opendatalab/MinerU2.5-Pro-2604-1.2B at the pinned revision d3f5e08d073c21466bbabe21c71bb1e9c2e595da; the converter (models/convert-mineru25-pro-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 cb0248e1eb545458fed0de1affefe5567ea1861d (workflow ocr-gguf-publish.yml, protocol docs/mineru25-pro-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 8 frozen requests (formula (29), table (140), receipt (202), german (389), japanese (37), arabic (7), layout-page (199), layout-rot90 (44) tokens), CrispEmbed commit 5c6005b9, reference revision d3f5e08d073c.
• F16 is token-ID exact on the same 8 cases (receipt tests/results/mineru25/precision.json); Q8_0 is token-ID exact on the same 8 cases (receipt tests/results/mineru25/precision.json).
• Not verified: GPU backends, multi-page documents, aggregate document benchmarks, contexts beyond the profile's 8,192 tokens.

Usage

MinerU2.5-Pro support is experimental and default-off in CrispEmbed:

CRISPEMBED_MINERU25_PRO=1 crispembed --ocr-pipeline page.png \\
  --ocr-engine mineru25-pro --ocr-rec mineru25-pro-2604-f16.gguf

Python (native inference with the pinned upstream document helper for full postprocessing):

from crispembed import CrispMinerU
with CrispMinerU("mineru25-pro-2604-f16.gguf", image_analysis=True) as m:
    print(m.to_markdown(m.two_step_extract("page.png")))