DeskNext 0.4.0 Preview 2 · 模型部署与默认模板修正

DeskNext 0.4.0 Preview 2 · Windows x64

English · 简体中文

Unsigned development preview. Back up important files and start with a test folder. No unattended organization is enabled. This is not P4/P7 completion or a production release.

Changes since Preview 1

• Default classification templates fit better. Opaque space GUIDs no longer consume the local model's limited option budget. Request-local short IDs map back to the same workspace spaces. Category descriptions and the two special choices are preserved; nothing is silently truncated. All 36 built-in template/language combinations passed the native-tokenizer check. This corrects input sizing, not proven classification accuracy.
• Model storage can move without losing the old cache. New downloads live under .desknext-downloads at the selected installation location. Complete verified previous downloads can be copied and reused offline; old files stay. Incomplete caches and retained model directories are not automatically deleted. Save settings after a successful install to activate it.
• Worker compatibility is explicit. Local requests and replies bind a protocol version and model-manifest digest. An incompatible host is refused rather than silently used. Native ONNX environment cleanup is explicit; the source-level macOS exit crash was reproduced and corrected with a native follow-up. This download still contains Windows x64 only.
• Windows type icons initialize COM on the background thread as required by the Shell API.

Download, run and upgrade

Download DeskNext-0.4.0-preview.2-win-x64.zip and extract the entire archive. Run payload/DeskNest.App.exe for a portable launch. Keep all payload files together. You do not need Python or an installed .NET SDK; the optional Laya model package remains separate and unchanged.

For a per-user install or upgrade, close DeskNext and follow the included README.md/Install.ps1. Do not overwrite an old version directory. Existing workspace GUIDs and model settings are not migrated by these changes. Uninstall preserves workspaces, models and unknown files. Follow your organization's execution policy; verify the checksum before deciding to run an unsigned build.

Checked scope

The actual installed executable passed the three frozen A–D reference comparisons, versioned CPU model loading, verified-cache reuse, explicit activation and reopened preview, default-template budget checks, and model suggestion → confirmation → journaled import → undo. Uninstall retained an unknown-file sentinel. The earlier A/B/C references were reused unchanged—not regenerated to match this build.

Still open: independent human-labeled quality data, live Jev acceptance, permanent model/cache removal, signing, clean-machine/hardware validation and remaining automation/release work. No automatic file moves, GPU validation or large-sample accuracy is claimed.

Source snapshot: 37423fa. Build: 0.4.0-preview.2. ZIP: 125,464,411 bytes, 450 verified payload manifest entries.

SHA-256 a3b0e546162b8803ea8ffe57ccbdce133ee671a758e508f7279e8cf304e0bab3

Project guide · Previous preview · Detailed evidence · Issues

---

简体中文

未签名开发预览版，不是正式发行或 P4/P7 全部验收完成。请备份重要数据，先用测试文件体验；无人值守自动整理仍未启用。

这次修正了什么

• 默认中文分类模板的输入预算问题：不再把冗长的空间 GUID 塞进模型选项。请求内的短 ID 仍映射到原来的空间，全部名称、说明和两个特殊选项保留，没有静默截断。十二语言、三种默认模板共 36 个组合通过原生分词检查；这不等于分类准确率验收。
• 下载缓存跟随安装位置：优先复制可用的完整已校验旧缓存，在新位置离线复用。旧文件保留，安装成功后仍需保存设置才能启用；不自动删除未完成缓存或旧模型。
• 推理进程校验版本和模型身份，不兼容时拒绝使用。源码中的 macOS 原生退出崩溃也已修正并定向复验；本次附件仍仅包含 Windows x64。
• 后台 Windows 文件类型图标补齐 COM 初始化，不改动真实用户文件。

启动与升级

下载下方 DeskNext-0.4.0-preview.2-win-x64.zip，完整解压后运行 payload/DeskNest.App.exe。不要只复制主程序；运行时无需 Python 或 .NET SDK。独立的模型包没有变化，不必因为更新应用而重复下载模型。

按用户安装/升级请先关闭 DeskNext，再按包内 README.md 运行 Install.ps1，不要覆盖旧版本目录。这些改动不迁移既有空间 GUID 或模型设置；卸载也会保留工作区、模型和未知文件。请遵守所在组织的脚本策略，并核对校验和后决定是否运行未签名程序。

此安装版通过了固定 A–D 对照、新协议 CPU 加载、缓存复用、保存启用与重开预览、默认模板预算、“建议→确认导入→撤销”以及卸载保留文件检查。旧 A/B/C 参考没有重新生成。人工质量数据、真实 Jev、永久清理、签名及剩余发行验收仍未完成，不以测试样例冒充全面可用性。

中文 README · 反馈问题；请不要上传私密文件、密钥或未脱敏的日志。

DeskNext 0.4.0 Preview 1 · Windows x64 / 开发预览

DeskNext 0.4.0 Preview 1 · Windows x64

A working desktop file workspace—with AI suggestions you remain in control of.

English · 简体中文

Unsigned development preview, not a production release. Back up important data and start with a small test folder. This release does not enable unattended file organization.

Download and start

Download DeskNext-0.4.0-preview.1-win-x64.zip below and extract the entire archive to a local folder. For a portable launch, open:

payload/DeskNest.App.exe

Keep the payload directory together; do not copy just the executable. The included runtime does not require an installed .NET SDK or Python. Models are separate and optional.

For a per-user installation, follow the included README.md and Install.ps1. The installer prints the installed application's path. Follow your organization's script policy; no system-wide policy changes are required. Windows may warn because this build is unsigned—verify the download against SHA256SUMS and review the source before deciding to run it.

What to try

• Create a managed space or map an existing folder; open a space in its own desktop window.
• Import a file or folder after reviewing its source and destination, then try the recorded Undo action.
• Use file menus, search, the collection capsule, tray access, list/grid/details views and app-only custom icons.
• Choose light/dark/system appearance, an accent color, or one of twelve interface languages.
• Download/install the optional local Laya model in Settings. Verification and a real CPU loading check run before a new model can be selected; save settings to activate it.
• Preview a classification, add a short clarification when a filename is ambiguous, and explicitly choose/confirm a destination. No model response moves files on its own.
• Optionally configure Jev with your own key and explicit sending consent. Your entered notes may be sent along with filenames, types and category descriptions; file contents and absolute paths are not automatically uploaded. API usage may be charged.

Recovery and data

Managed deletion moves items to the application's recovery area; removing a mapped reference does not delete its original. Undo refuses changed or insufficiently identified items rather than overwriting them. Outgoing drags request Copy and never delete their sources. Model repair installs a fresh copy and leaves the old directory intact. Application uninstall preserves workspaces, model caches and unknown files.

Scope of this build

This Windows x64 snapshot was installed and exercised with the real CPU worker. Three fixed A–D comparison cases matched preprocessing, rankings and actions within the documented numerical tolerances. A separate installed-app fixture completed model suggestion → explicit confirmation → file import → undo, and uninstall retained its unknown-file sentinel.

That is not a large-sample classification-quality result, GPU validation, clean-machine certification, power-loss guarantee or signing/notarization acceptance. Local previews currently cold-start a worker. Live Jev evaluation, model removal/cache maintenance, broader automation/Flow and remaining release work are still open. macOS/Linux source builds have separate evidence; this asset contains only Windows x64 binaries.

Build identity: 0.4.0-p4-parity-20261005, application source snapshot 936e102. The archive is renamed for readability; its bytes are unchanged from the checked development bundle.

SHA-256:

2ed4a34b729fd9a619da178211fd47d105f5ec8651efd7a1aa52971749a14b01

Project / build guide · Model package · Detailed coverage · Report an issue

---

简体中文

可以开始体验的桌面文件工作区：AI 提供建议，每一次移动仍由你确认。

这是未签名的开发预览版，不是正式发行版。请备份重要数据，先用少量测试文件体验。当前不启用无人值守自动整理。

下载与启动

下载下方 DeskNext-0.4.0-preview.1-win-x64.zip，完整解压到本地文件夹，打开 payload/DeskNest.App.exe 即可便携运行。请保留整个 payload 目录，不要只复制主程序。运行时不需要安装 .NET SDK 或 Python，模型独立下载且不是手动整理的前提。

如需按用户安装，请按压缩包内的 README.md 运行 Install.ps1，安装器会给出启动路径。遵守所在组织的脚本策略，不需要修改系统全局策略。由于未签名，Windows 可能提示风险；请核对 SHA256SUMS 并审阅来源后自行决定是否运行。

建议体验顺序

1. 新建托管空间或映射已有文件夹，再打开独立空间窗。
2. 拖入文件/目录，核对来源和目标后确认导入，再体验操作历史中的撤销。
3. 使用搜索、收集箱、托盘、排列方式、自定义图标与主题色；界面支持十二种语言。
4. 在设置中下载或离线安装 Laya 模型，等待校验与 CPU 加载检查，保存设置后启用。
5. 请求分类建议；文件名不清楚时补充简短说明，再手动选择并确认目标。模型本身不会执行文件移动。
6. 如需 Jev，可填写自己的密钥并明确同意发送。可选补充说明会与文件名、类型、分类说明一起发送；不自动读取或上传正文和绝对路径，API 可能收费。

数据与限制

托管删除进入应用恢复区，移除映射引用不删除原文件；身份不足或内容变化时撤销会拒绝覆盖。外拖只请求复制，不删除源文件。模型修复生成新副本，保留旧目录；卸载应用也会保留工作区、模型缓存和未知文件。

本包已通过安装版 CPU worker 的三个固定案例 A–D 对照，以及独立的“建议 → 明确确认 → 导入 → 撤销”文件流程和卸载保留文件检查。这不代表大样本分类质量、GPU、干净机、掉电耐久性或签名验收。Jev 实际评测、模型清理、自动化/Flow 和其余发行工作仍在推进。当前附件仅供 Windows x64；macOS/Linux 的源码与证据分别维护。

构建标识为 0.4.0-p4-parity-20261005，源码快照与 SHA-256 见上方。详情见中文 README，问题请提交到 Issues，不要附带私人文件、密钥或未脱敏的恢复记录。

Laya Multilingual FP32 — Local Model Preview / 本地模型预发布

Laya multilingual FP32 — DeskNext model preview

English · 简体中文

This release contains the optional local classification model, not the DeskNext application. It is a fixed multilingual FP32 ONNX package for the app's Laya CPU preview. No Python environment or cloud key is required at runtime.

Download

Asset	Purpose
desknext-laya-multilingual-fp32-v1.zip	Complete model bundle, 812,781,615 bytes (about 813 MB / 775 MiB)
SHA256SUMS	Archive integrity checksum

bd2464a8b63f195fd1aed579b355b37d3ef6f45f1e796116b041846a7de3f6fc

In the current DeskNext source build, use Settings → Laya → Download model for the cancellable online path, or Install model package to select this ZIP offline. Wait for verification and installation, then Save settings to activate the resulting model folder. Alternatively, extract the complete package and select its directory; keep both .onnx.data files beside their graphs. Model installation does not move your classification files.

Included and traceable

• Split encoder.onnx and head.onnx graphs plus external data, tokenizer and calibration configuration; ONNX opset 18.
• Pinned Laya source: 970dc8c5f63d7b886a68409493f37d569424f933.
• Multilingual checkpoint: 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851.
• Ten archive entries, including the original Apache-2.0 license, modification notice and export patch.
• The application pins archive and manifest hashes and verifies model files before activation.

The package contains no user files, credentials, application binaries or training dataset. Earlier fixed-input CPU equivalence and local preview checks are documented in the repository; they do not establish general classification accuracy, GPU support, all-platform compatibility or automatic-move safety. This remains a model prerelease, not a signed application release. A checksum pins bytes; it is not a publisher signature.

See the project overview, 中文说明 and deployment status.

---

简体中文

本预发布包含可选的本地分类模型，不是 DeskNext 应用安装包。它是供 Laya CPU 预览使用的固定多语言 FP32 ONNX 模型，运行时无需 Python 或云端密钥。

下载与使用

• desknext-laya-multilingual-fp32-v1.zip：完整模型包，812,781,615 字节，约 813 MB / 775 MiB。
• SHA256SUMS：压缩包完整性校验文件，SHA-256 见上方。
• 在当前 DeskNext 源码构建中，打开设置 → Laya → 下载模型使用可取消的在线安装，或通过安装模型包离线选择 ZIP。等待校验与安装后，再点击保存设置启用新模型目录。也可完整解压后选择目录；不要拆散图文件与对应的 .onnx.data 文件。
• 模型安装不会移动你的分类文件，也不会隐式授权自动整理。

来源与边界

模型包包含 encoder/head 双图、外部数据、分词器、校准配置，以及 Apache-2.0 许可证、修改声明和导出补丁；共十个文件。固定上游源码和 checkpoint 版本见上方。应用会校验压缩包、清单及模型文件。

包内没有用户文件、密钥、应用二进制或训练数据。既有 CPU 固定输入等价检查与本地预览证据，不等于分类质量、GPU、所有平台或自动移动验收。本条目仍为模型预发布，不是已签名的应用正式版；校验和也不等于发布者签名。