Refresh the bundled MNN runtime with the latest CPU LinearAttention and Arm82 fp16 optimization path. Improve tokenizer and template rendering compatibility for thinking-mode prompts and array concatenation cases. Add TopKV2 backend coverage for OpenCL and Metal execution paths. Bugfix: Fix the crash when tapping Add Local Model on Android. Avoid deadlocks caused by partially initialized mmap weights during model loading. Reuse the loaded runtime session when starting API service for the same model to avoid extra reloads that could freeze or crash API startup.

mnn chat是一款开源多模态大型模型工具，它能在移动设备上自主运行复杂模型，无需依赖云端服务器，此应用全面适配多种模型及多模态任务，有力保障了用户数据的安全与隐私权益，无论是撰写文字还是识别图像，可以在不同品牌不同操作系统的手机上面运行，流畅不卡顿而且引入的大模型的占存不大，软件会将其压缩量化，为手机等设备提供流畅的AI对话与模型推理体验。