Refresh DotsOCR mmproj with corrected converter output

README.md

@@ -25,6 +25,14 @@ GGUF conversions of [rednote-hilab/dots.ocr](https://huggingface.co/rednote-hila
 | Dots.Ocr-1.8B-F16.gguf | 3.4 GB | Text model, float16 |
 | mmproj-Dots.Ocr-F16.gguf | 2.4 GB | Vision encoder (mmproj), float16 |
 
+## Update
+
+On March 23, 2026, `mmproj-Dots.Ocr-F16.gguf` was regenerated from a corrected DotsOCR converter. The text GGUF files did not change. If you downloaded the `mmproj` earlier, refresh that file.
+
+Current llama.cpp fork with DotsOCR support and the compatibility fix:
+
+- [anthony-maio/llama.cpp](https://github.com/anthony-maio/llama.cpp)
+
 ## Architecture
 
 dots.ocr = Qwen2 text backbone (1.7B params, 28 layers) + modified Qwen2-VL vision encoder (1.2B params, 42 layers).
@@ -36,6 +44,34 @@ Key differences from Qwen2-VL:
 
 ## Usage with llama.cpp
 
+Requires a llama.cpp build with DotsOCR support. At the moment, use:
+
+- [anthony-maio/llama.cpp](https://github.com/anthony-maio/llama.cpp)
+
+Single-image example on Windows:
+
+```powershell
+llama-mtmd-cli.exe `
+  -m .\Dots.Ocr-1.8B-Q8_0.gguf `
+  --mmproj .\mmproj-Dots.Ocr-F16.gguf `
+  --image .\page.png `
+  -p "Extract all text from this image and preserve structure in markdown." `
+  --ctx-size 131072 `
+  -n 4096 `
+  --temp 0 `
+  --jinja
+```
 
+Equivalent server launch:
 
-> **Note:** Requires llama.cpp with dots.ocr support (pending upstream merge: https://github.com/ggml-org/llama.cpp/pull/19882/changes)
+```powershell
+llama-server.exe `
+  -m .\Dots.Ocr-1.8B-Q8_0.gguf `
+  --mmproj .\mmproj-Dots.Ocr-F16.gguf `
+  --port 8111 `
+  --host 0.0.0.0 `
+  --ctx-size 131072 `
+  -n 4096 `
+  --temp 0 `
+  --jinja
+```

mmproj-Dots.Ocr-F16.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ea68c0dfa6d23eb07a5f2cf91e84d65f55cf26037088b430df8850597565618
-size 2524495776
+oid sha256:b65a1db58db75f6d1ae900a4cb80772c4c0fabfaf261439e4c4ab965f6970de8
+size 2524495808