llama.cpp b10287 fixes Unlimited-OCR max_tiles converter setting
llama.cpp's new release fixes OCR limits with a max_tiles converter option.
ggml-org has shipped llama.cpp b10287, a fresh release of the popular C/C++ inference engine for LLMs. The standout change is a fix for Unlimited-OCR, specifically adding a max_tiles setting to the conversion script. This addresses a previous limitation where OCR processing would fail or truncate on documents with many tiles, making the tool more reliable for large-scale or dense document extraction.
The release was tagged on August 5 and comes with prebuilt binaries across nearly every major platform. macOS users get Apple Silicon builds (both standard and KleidiAI-accelerated), Intel x64, and iOS XCFramework. Linux covers x64, arm64, s390x, plus GPU backends including Vulkan, ROCm 7.2, OpenVINO, and SYCL in FP32/FP16 variants. Windows users get CPU, ARM64, OpenCL Adreno, CUDA 12.4/13, Vulkan, OpenVINO, SYCL, and HIP builds. Android and openEuler builds are also included. This breadth means developers can adopt the OCR fix immediately on their preferred stack, from local laptops to cloud GPU instances.
- Fixes Unlimited-OCR by adding a max_tiles setting to the converter script
- Release b10287 on Aug 5 from ggml-org, signed with GitHub verified signature
- Prebuilt binaries for macOS, Linux, Windows, Android, and openEuler with CPU, Vulkan, CUDA, ROCm, and more
Why It Matters
For teams running local LLM pipelines, this fixes OCR truncation on large docs, enabling reliable document processing at scale.