Skip to content

Add PaddleOCR-VL 1.6 to the llama.cpp OCR model list#12371

Merged
niksedk merged 1 commit into
mainfrom
feature/llamacpp-paddleocr-vl
Jul 11, 2026
Merged

Add PaddleOCR-VL 1.6 to the llama.cpp OCR model list#12371
niksedk merged 1 commit into
mainfrom
feature/llamacpp-paddleocr-vl

Conversation

@niksedk

@niksedk niksedk commented Jul 11, 2026

Copy link
Copy Markdown
Member

Adds PaddlePaddle's official llama.cpp package of PaddleOCR-VL 1.6 (NaViT + ERNIE-4.5, 109 languages; model + mmproj, 1.8 GB total) as a third entry in the llama.cpp OCR model list, alongside GLM-OCR and LightOnOCR. One list entry — the existing download/server/OCR plumbing handles the rest.

This restores the 109-language PaddleOCR-VL option that was evaluated and dropped from the CrispEmbed backend list in #12366 (broken there in CrispEmbed v0.14.0's integration — this verification confirms the model family itself is fine when run through llama.cpp).

Verification (headless, pinned llama.cpp b9833, Apple M4)

Driven through SE's actual LlamaCppServerManager.EnsureServerRunningAsync + LlamaCppOcr code path:

  • Real Blu-ray sup (alien4.sup, 6 lines sampled across the movie): 6/6 exact, 0.3–4.3 s/line
  • Synthetic subtitle image ("Hello world, this is a test.", white on black): exact — notably the same image CrispEmbed's Paddle path truncated
  • Multi-line Danish/Spanish/French with accents and italics: line break and text exact, one diacritic slip (Àh for Åh)

Full UI test suite passes (608 tests).

🤖 Generated with Claude Code

PaddlePaddle's official llama.cpp package (model + mmproj, 1.8 GB
total) - NaViT + ERNIE-4.5, 109 languages. Verified headlessly against
the pinned llama.cpp b9833 via SE's LlamaCppServerManager/LlamaCppOcr
path: 6/6 exact lines on a real Blu-ray sup (0.3-4.3 s/line on an M4)
and correct accented multi-line output on synthetic subtitle images.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@niksedk niksedk merged commit 7ca7c37 into main Jul 11, 2026
1 of 3 checks passed
@niksedk niksedk deleted the feature/llamacpp-paddleocr-vl branch July 11, 2026 19:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant