OvisOCR2 for CrispEmbed (f16)

This GGUF contains the complete main-model checkpoint of ATH-MaaS/OvisOCR2 at cbdede82e048659b83310476ff807da825f8fe78, converted for the experimental CrispEmbed format-2 OvisOCR2 reader. The pinned upstream card and LICENSE declare Apache-2.0. The audited 473-tensor checkpoint is the non-speculative main-model schema. The source config declares MTP1, but this export makes no MTP support claim. Stock llama.cpp compatibility is not claimed.

The Ovis card declares Qwen3.5-0.8B as its base; that card declares Qwen3.5-0.8B-Base. All three pinned metadata records declare Apache-2.0. Their LICENSE text is identical after newline normalization; the root Base file has a different raw SHA-256 due to line endings. The exact chain, raw file hashes and inspection scope are in manifest.json. No separate NOTICE file was listed at those pins.

The pinned upstream card is also preserved as UPSTREAM_README.md. The audited CrispEmbed source commit is b33c5bad6d47433ea1f3f1fa1c9025d3d22e6fe5. This is a CPU-only experimental reader. After the original image resize it accepts at most 1024 visual patches; the original 2880²-pixel maximum has not been qualified. Until a qualified release includes it, build that source and enable the three experimental entry gates and the audited vision math settings to load the GGUF by its explicit local path:

CRISPEMBED_QWEN35_OCR=1 CRISPEMBED_OVISOCR2_FRONTEND=1 CRISPEMBED_OVISOCR2_INFERENCE=1 CRISPEMBED_QWEN35_VISION_PRECISE_NORM=1 CRISPEMBED_QWEN35_VISION_PRECISE_ATTN=1 CRISPEMBED_QWEN35_VISION_CASCADE_NORM=0 \
  crispembed -t 2 --ocr-pipeline page.png --ocr-engine ovisocr2 \
    --ocr-rec ovisocr2-f16.gguf --ocr-max-tokens 1024

Recognition returns raw generated Markdown/HTML. The quality reports separately record the upstream parser cleanup/filter view. The reader uses the pinned original task prompt and rejects a custom prompt on the explicit Ovis stage. The bounded original/native, public API, independent quality and quantization-selection receipts were reviewed before this file was uploaded. This candidate uses no importance-matrix calibration. Their exact SHA-256 values and source commits are in manifest.json. The quality panels cover four short self-authored heldout pages and three reviewed documents; they do not establish broad OCR accuracy. The publication workflow verifies the uploaded GGUF again from an empty cache at its immutable Hugging Face commit.

The immutable CrispEmbed evidence revision is 121dc94349a247101ecfb0b038cefe4b8d209930:

• Native stages/cache/logits/IDs
• C/Python/CLI/HTTP/Rust/Dart interfaces
• Original Python reference and first F16 quality
• Selected f16 quality and all quantization outcomes/selection
• Measured quality, timing boundaries and limits

The first F16 run matches all1841 generated Python IDs and all eleven raw texts, with EOS and no caps. The original receipt/two-column recognition errors remain. Vision retains four raw strict0.002 exceedances while passing its pre-existing original-backend calibrated budgets; the linked reports retain both views. The CPU reader has substantial latency: the first F16 receipt took about23 minutes on an EPYC7763 with two inference threads. Its HF comparison used a different CPU and timing boundary. See the linked measurements before choosing this experimental reader for interactive work.

No default model alias, CrispEmbed runtime release, or app selection is activated by this model repository. Original model probes and test-only weights are held separately from this distribution.