v5.1.1

Desktop Client

Platform	Architecture	Binary
Linux	x64	mtranserver-desktop-5.1.1-linux-x86_64.AppImage
Linux	arm64	mtranserver-desktop-5.1.1-linux-arm64.AppImage
Linux	x64 (deb)	mtranserver-desktop-5.1.1-linux-amd64.deb
Linux	arm64 (deb)	mtranserver-desktop-5.1.1-linux-arm64.deb
Linux	x64 (rpm)	mtranserver-desktop-5.1.1-linux-x86_64.rpm
Linux	arm64 (rpm)	mtranserver-desktop-5.1.1-linux-aarch64.rpm
macOS	x64	mtranserver-desktop-5.1.1-mac-x64.dmg
macOS	arm64	mtranserver-desktop-5.1.1-mac-arm64.dmg
Windows	x64	mtranserver-desktop-5.1.1-win-x64.exe

桌面端启动后会创建一个托盘菜单，通过菜单可以方便的管理服务。

macOS: 拖动安装后，如果打开 MTranServer App 提示有问题无法打开，需要运行以下命令修复：

macOS: If you see any error when opening MTranServer App after installing, run this command in Terminal App:

sudo xattr -d com.apple.quarantine "/Applications/MTranServer.app"

Server

v5 服务端请使用 Docker 或通过命令行安装。暂不提供服务端单文件可执行程序。

For v5 server use, choose Docker or install from the command line. Standalone server executables are not currently available.

npx mtranserver@latest

Docker image: xxnuo/mtranserver:latest

Upgrading and Offline Use

升级时请保留配置和翻译模型文件夹。首次使用或升级后，程序可能需要下载并准备翻译文件，请保持联网并等待完成。断网前先试译需要使用的每个翻译方向。

Keep your configuration and translation model folders when upgrading. First use or an upgrade may require downloading and preparing translation files. Stay online and wait for preparation to finish. Try every required translation direction before going offline.

Changelog

v5.1.1

• 修复自动调优因探测输入达到生成上限而误报“没有可用 ONNX 后端”的问题；使用真实输入片段测速，实际翻译仍严格检查输出是否完整。
• Docker 镜像补充 CA 证书和 Vulkan/Mesa 依赖，支持容器内 HTTPS 证书校验及 WebGPU 探测。
• 提供默认纯 CPU、Intel、AMD 和 NVIDIA 四份 Compose 配置；默认跳过 GPU 探测，GPU 配置继续按实测性能选择后端并保留 CPU 回退。
• Intel/AMD 配置映射 /dev/dri；NVIDIA 配置通过 NVIDIA Container Toolkit 使用宿主驱动，以 WebGPU/Vulkan 推理，不包含 CUDA/TensorRT 依赖。

English

• Fix automatic tuning incorrectly reporting no usable ONNX backend when a timing probe reaches the generation limit. Benchmark real input fragments while still requiring complete output for actual translations.
• Include CA certificates and Vulkan/Mesa dependencies in Docker images for HTTPS certificate verification and WebGPU discovery.
• Provide separate CPU-only, Intel, AMD and NVIDIA Compose configurations. The default skips GPU discovery; GPU configurations retain measured backend selection and CPU fallback.
• Map /dev/dri for Intel/AMD. The NVIDIA configuration uses host drivers through NVIDIA Container Toolkit for WebGPU/Vulkan inference; CUDA/TensorRT dependencies are not included.

v5.1.0

• 根据真实推理耗时自动选择可用的 CPU/GPU 后端和 CPU 线程配置，不再固定使用单线程。
• 启用完整图优化、内存复用和线程自旋，实测执行模式、负载均衡及有上限的 CPU 并发，保留更快的配置。
• 同一进程内复用模型的调优结果；加速后端运行失败时自动重试 CPU，并等待已接收的请求完成后释放引擎。
• 补充关闭缓存的 ONNX 性能汇总表和版本对比。

English

• Automatically select available CPU/GPU backends and CPU thread counts using real inference measurements instead of a fixed single thread.
• Enable full graph optimization, memory reuse and thread spinning; measure execution modes, load balancing and bounded CPU concurrency to retain faster settings.
• Reuse model tuning within the process, retry failed accelerators on CPU, and drain accepted requests before releasing an engine.
• Include summary tables of uncached ONNX performance and version comparisons.

v5.0.0

• 更换翻译引擎，首次使用或升级后可能需要重新准备翻译文件，请耐心等待。
• 下载翻译文件时会自动完成准备，之后可重复使用。
• 离线使用前，请先联网试译需要使用的每个翻译方向；迁移时保留配置和翻译模型两个文件夹。
• 服务端使用 Docker 或命令行安装，暂不提供单文件可执行程序。
• 更新使用说明和界面提示，修正设置说明，移除无效的每语言工作数设置和旧版性能数据。

English

• Updated the translation engine. First use or an upgrade may require preparing translation files again; please wait for this to finish.
• Downloading translation files now also prepares them for reuse.
• Try each required translation direction online before going offline. Keep both configuration and translation model folders when moving files.
• Use Docker or a command-line installation for the server. Standalone server executables are not currently available.
• Updated guides and interface messages, corrected settings descriptions, and removed the ineffective workers-per-language setting and old performance figures.