中文
- 简化自动档位推荐：移除推荐流程中的显存检测与容量门槛，直接复用小模型后端实际选择的 CPU/GPU 设备。选中 GPU 时开放全部四档，默认使用 Standard，Advanced 需要显式指定；选中 CPU 时，按进程可用逻辑核数推荐：不足 4 核仅开放 Flash，4–11 核开放到 Basic，至少 12 核开放到 Standard。
- 统一 llama.cpp 运行策略：Linux 和 Windows 自动模式复用小模型后端的 CPU/GPU 选择，具体显卡由 llama.cpp 选择。这些平台上的 GPU 主模型默认使用 4 个生成线程，保留用户显式设置。
- 改善 PDF 文本清理：在后置 OCR 和失败恢复完成后，将原生正文中残留的 Unicode 私用区字符替换为空格，减少异常字体映射造成的乱码，覆盖基本和补充私用区。
English
- Simplified automatic tier recommendations: Remove GPU memory detection and capacity thresholds from tier recommendations, reusing the CPU/GPU device actually selected by the small-model backend. GPU selection enables all four tiers, with Standard as the default and Advanced available by explicit request. CPU selection uses the logical cores available to the process: fewer than 4 enables Flash only, 4–11 enables up to Basic, and 12 or more enables up to Standard.
- Unified llama.cpp runtime policy: Automatic mode on Linux and Windows reuses the small-model backend's CPU/GPU selection and lets llama.cpp choose the specific GPU. On these platforms, GPU models default to 4 generation threads while preserving explicit user settings.
- Improved PDF text cleanup: Replace residual Unicode private-use characters in native body text with spaces after post-OCR processing and fallback restoration, reducing garbled text caused by unusual font mappings. Both basic and supplementary private-use areas are covered.
Full Changelog: mineru-4.1.0-released...mineru-4.1.1-released