The llama.cpp project released version b9963, which introduces support for DeepSeek OCR v1 with multi-tile dynamic resolution. This update includes unified image preprocessors for both DeepSeek OCR versions and fixes for handling multi-row and multi-column images.

  • Added deepseek-ocr v1 multi-tile dynamic resolution and unified image preprocessors.
  • Fixed ds-ocr img-preproc fuse_row tile-drop issues for multi rows and columns.
  • Removed duplicate redundant img_end and simplified DeepSeek OCR graph CLS broadcast cleanup.
  • Relaxed test tolerance and made DRY opt-in with model-specific n_predict.

The release provides binaries for macOS, Linux, Android, Windows, and openEuler across CPU, GPU (CUDA, ROCm, Vulkan, OpenCL), and AI accelerator backends.