Hex 3.4.02026-11-11
Hex 3.4.0
Windows 用户从 3.3 升级时若只提示下载更新,直接运行安装包覆盖安装即可,设置、历史与模型都会保留。Hyprland 用户需注意:旧的带 key 名称的绑定仍可用,但录制新按键前请改用修正后的名称(anomaly.hex:dictation、paste-last、capture、raw-dictation),重载配置后用 hyprctl globalshortcuts 核对。
本次新增:首次运行改为连贯的引导故事;界面新增俄语;新增 orukeet、granite speech 5.0 470m turboctc 及挪威语 nb-whisper medium,能自动识别语言的模型归入 auto 分组。历史记录头部整合为「保留 7 天 ▾」菜单,每条记录标注所用模型与语言,重新转写可保留最多五个旧版本。会议功能(需手动开启)的姓名与摘要选项按平台区分,转写不会默认外发给 opencode;会议行显示通话应用图标,Discord 通话也能被识别。听写模式新增独立取消快捷键(默认 Escape),「听写时」静音、暂停或压低其他音频不再有预滚。Linux 与 Windows 滚动条常驻可拖动,菜单支持翻页与 Home/End;触屏点击、拖拽与惯性滚动可用;窗口、HUD 与胶囊均补充了无障碍名称与状态。
Windows 端 hex.exe 不再从开始菜单弹出控制台,POST /dictations 支持自带麦克风并可流式输出文本。macOS 端菜单栏重新排序,密码框等安全输入场景下快捷键依然生效。
实验室功能(默认关闭)包括:mac 截图标注与 HEVC 录屏、刘海指示器、Caps Lock 免手动听写、每模式独立的语音模型与语言、历史录音回放、Wayland 实时输入等。
修复方面,mac 听写中途麦克风失联会自动重开而非丢弃,中断的听写会被保留并复制;HUD 重新加入所有空间,底部位置与刘海场景正常,实时文本不再闪烁上一条内容。
Hex 3.3.02026-11-02
Hex 3.3.0 更新
首次运行时,演示与练习环节可直接跳过;语音模型下载失败或卡住也不再阻塞流程,可稍后自行选择。新增模型选择器,支持三款模型并排对比,也能浏览全部可用模型。俄语新增 gigaam v3,支持导入自定义 gigaam 模型文件。界面新增繁體中文。快捷键模式收进「通用」设置,Hex 只会列出当前桌面支持的模式,且每种模式可配置独立词表。安装新版时,仍在运行的旧版本会自动退出。macOS 支持通过配置描述文件锁定功能开关;Linux 新增本地 API,第三方应用可借此启停听写,与 macOS 对齐。实验室新增三项:转写前跳过长停顿、通话结束 30 秒后自动停止会议录音、在输出框直接修正错误结果。
修复方面:macOS 快捷键在 cmd+tab、cmd+q 残留按键状态或闲置后失效的问题已解决,fn 键也可用;Windows 修复麦克风卡在「启用中…」无法听写,以及 Stream Deck 等注入式按键被吞掉的问题。剪贴板格式报错、口述列表与数字被自动清理改写、历史记录清除后回滚等问题一并修复。opencode 2 版本过旧时会明确提示升级,不再静默跳过 AI 改写。Linux 修复外部修改设置不生效、状态栏不显示麦克风信息、GNOME + NVIDIA 下 Xwayland 启动崩溃;KDE 快捷键默认改为 ctrl+alt+space 并需用户确认,GNOME 注册更可靠,niri 等无法绑定快捷键的桌面可通过 hex dictation 配合自定义绑定使用。annotate 修复 Hyprland、分数缩放、Intel + NVIDIA 笔记本录制等问题;GPU 预热失败的自定义模型可重新加载,语音引擎不再误报 GPU 无响应,会议界面的就绪与录制胶囊新增关闭按钮。
Hex 3.2.02026-11-01
Hex 3.2.0 带来全新的首次运行流程:无论哪个系统,安装后都会自动配置环境并下载对应语言的语音模型,随后用几条练习消息直接上手快捷键。若检测到电脑里已有 Hugging Face 缓存的语音模型,会与推荐模型并列展示,可跳过下载。新增希伯来语(ivrit.ai 的 whisper large v3 turbo)与 nemotron 3.5 asr streaming 0.6b 两种模型,界面新增巴西葡萄牙语。麦克风改为选择器旁的单一开关,设置中注明各模式对隐私的影响;语音引擎新增 auto / GPU / CPU 选项,遇到卡死过的 GPU 会自动回退到 CPU。
修复方面:带直流偏置的麦克风(如部分笔记本内置麦)听写不再在 2.5 秒后被截断(#841);Escape 不再取消听写,误触不会丢掉已说内容。Windows 下听写不再堆积进剪贴板历史(Win+V),远程桌面粘贴恢复正常;Linux 修复 KDE Wayland 热键、快捷键录制键位与 Hyprland 文档,Intel+NVIDIA 笔记本可录制视频;macOS 修正副屏录音指示位置,并识别 gather 通话。长菜单会提示可滚动,官网加载也更快了。
Hex 3.1.22026-11-01
Hex 3.1.2
Fixes
- Dictation no longer gets cut off. With a quiet built-in microphone, Hex
could keep only the first 2.5 seconds of a dictation and paste that as if
nothing was wrong. It now keeps all of it, and a dictation that comes out
empty is never dropped silently.
- Escape no longer cancels a dictation, so a stray key press can't throw
away what you said.
- Windows: numpad keys work as shortcuts with Num Lock off, punctuation
shortcuts keep working after a keyboard layout switch, and the shortcut
recorder shows your layout's punctuation.
- History: clearing History stays cleared. A background cleanup could write
an old copy back, restoring cleared entries and dropping new ones.
- Meetings: calls in Arc (including Google Meet) and WhatsApp are detected.
- The recording indicator follows the pointer to another display during a
dictation.
- Linux: Hex runs in the background and launches at login through the
desktop's Background portal, GNOME without a shortcuts portal can add
Hex's shortcuts in one click, and quitting works on NVIDIA Wayland.
- macOS: Korean and other CJK text displays correctly.
New
- Hebrew, with ivrit.ai's fine-tune of Whisper Large v3 Turbo.
- Nemotron 3.5 ASR Streaming 0.6B.
- Hex can use speech models already on your computer instead of downloading
them again.
Hex 3.1.12026-11-01
Hex 3.1.1
Fixes
- Hex opens again when a setting it can't read is in its settings file. 3.1.0
quit silently at launch for anyone who had used a recent Dev build.
- GNOME and KDE (Wayland): asking for and approving global shortcuts works,
with a safe default shortcut, a warning when it clashes with the desktop's,
and a hint for asking again if the desktop declines. Paste closes its
remote-control session after each paste, including on GNOME 46, and KDE
restores your clipboard afterwards.
- Linux: a long dictation no longer sits in Transcribing for a minute after
the OpenCode rewrite times out. Hex's sound output closes after its tones,
so it doesn't linger in the desktop's sound controls. Launch at login keeps
working after Hex moves, and the window resizes properly on floating
desktops.
- Whisper Large v3 Turbo accepts Tamil and Telugu.
- macOS: an external microphone works with the lid closed, and Settings clears
stale permission rows with one-click recovery.
- Meetings are kept for 24 hours by default; older meetings stay until you
choose how long to keep them.
- The homepage at https://anoma.ly/hex is faster.
New in Labs
Each is off until you turn it on in Settings › Labs:
- Left and right modifier keys as separate shortcut keys.
- Hold and Toggle together as a dictation shortcut mode.
- Filters and sorting in the speech model picker.
- Benchmark explanations in the speech model picker.
Hex 3.1.02026-11-01
Hex 3.1.0
Requirements
- On Windows and Linux x64, Hex now needs a 2013 or newer processor with AVX2
(Intel Haswell / AMD Excavator or later). On an older processor Hex shows a
message and quits instead of crashing.
Fixes
- Windows: the recording indicator appears reliably, and says why when it
can't. Laptops with both an integrated and a dedicated GPU no longer crash
at launch, and transcription survives a GPU that crashes.
- Windows: paste works in more apps, including VS Code's terminal.
- Meetings: starting, saving and deleting recordings, and the error banner,
are fixed.
- Annotate captures the screen and draws correctly again.
- Linux: global shortcuts work on GNOME and KDE Wayland, including changing
and cancelling them, and Hex's icon shows everywhere.
- The recording indicator follows you: it appears on the monitor under the
pointer.
- Hex only talks to an OpenCode that's already running and no longer starts
one itself.
- Fixed a crash at startup on some Windows and Linux PCs (the CPU fix above).
- Dictation with a Bluetooth microphone, double-tap on X11, and punctuation at
the end of a dictation are fixed.
New
- QwenCleo-ASR 1.7B for Egyptian Arabic.
- Recognition hints now apply to Qwen3-ASR models, and hints are merged into
Vocabulary, so there's one list of names and terms.
- Meetings can be kept for 24 hours, 7 days, 30 days, or forever.
- New Labs toggles, each off until you turn it on: free the speech model's
memory when unused, spoken formatting ("new line", "new paragraph",
quotes), remove filler words, and Restore default for the Paste last
dictation shortcut. On macOS also: keep the dictation on the clipboard,
copy instead of paste, a second dictation shortcut, smart spacing, and
subtitle export for meetings.
Moved to Labs (now off until you turn them on)
Turn these back on in Settings › Labs:
- macOS: mouse buttons as shortcuts.
- Switching the speech model with the keyboard language.
- macOS: pausing the browser along with other media while dictating.
- Windows: letting apps such as Discord see the dictation shortcut too.
Hex 3.0.02026-10-31
Hex 3.0.0
Upgrading from 2.3.1
Hex 3.0 starts fresh. It keeps its data in a folder named hex instead of
voice-control and imports nothing, so settings, History, meeting recordings,
and downloaded models start empty and the first launch downloads a speech
model again. Meeting notes already exported to ~/Documents/Hex Meetings stay
where they are. The old voice-control folder is left alone, and you can
delete it.
2.3.1 doesn't update to 3.0 by itself. Download Hex from
https://anoma.ly/hex.
- macOS: Hex is now anomaly.hex, signed by Anomaly, so macOS treats it as a
new app. Remove the old Hex entries in System Settings › Privacy & Security,
then allow Microphone, Accessibility, and, for Meetings, Screen & System
Audio Recording again. Turn Launch at login back on if you used it.
Configuration profiles that set DisableUpdates need to target
anomaly.hex.
- Windows: the installer is a new app in Settings › Apps, so installing 3.0
leaves 2.3.1 installed. Uninstall the old Hex by hand.
- Linux: install again with the command on https://anoma.ly/hex.
Removed
- Voice Action is gone.
- Spoken commands are gone, along with the Moonshine model they used.
hex.config.ts now holds only transformations.
New
- A new settings window, the same on macOS, Windows, and Linux. The sidebar
has General, Output (replacements, per-app modes, OpenCode rewriting, and
transformations), and History. On macOS, Labs holds Meetings (beta), which
joins the sidebar once turned on; Dev builds add Annotate and experiments.
- A new recording indicator on every OS.
- Vocabulary (Settings › General, or hex vocabulary): list names and terms,
and Hex steers transcription toward them and fixes common mishearings.
- Shortcuts can use both Option keys together, a left or right modifier, or a
mouse button.
- A new set of feedback sounds, with the old ones under Classic, and a picker
you can listen to.
- On macOS, Hex can lower other audio to a volume you choose while you
dictate, and pausing media now pauses the browser too.
- Local models use about 160 MB of memory instead of about 900 MB.
- New speech models, Underdog Bark 3.0 Pro and Clean, and Bangla. The picker
shows each model's accuracy.
- Ukrainian interface language.
- Hex skips recordings with no speech in them instead of transcribing them.
- macOS no longer asks for Input Monitoring. Accessibility covers the
shortcut.
- Meetings (beta, in Labs on macOS) can remind you before calendar meetings
once you connect a calendar, can always record calls from an app you choose,
and finishes a meeting that was recording when Hex quit on the next launch.
- Annotate (Dev channel, macOS and Linux, in Labs) records the screen while you
draw and talk, and sends it to an OpenCode session.
- The hex command controls the running app (hex status, hex dictation,
hex settings, hex meetings record) and reads History and meetings
(hex history, hex meetings, hex mcp).
- One updater on every OS, with a Stable or Dev update channel in Settings ›
General.
- Linux gets ARM64 builds, .deb and .rpm packages, and a tray icon.
This macOS release uses build 30000.
Hex 2.3.12026-10-21
> Linux Wayland users: don't add yourself to the input group to make 2.3.1's shortcuts work. Install the new build, which uses the desktop's GlobalShortcuts portal or hex commands bound in your compositor instead: see Upgrading from 2.3.1.
Hex 2.3.1
- Plays the start sound as soon as you press the dictation shortcut, instead
of after you've held it.
- Offers to record calls in Slack and other Chromium-based apps, whose audio
runs in a helper process.
- Shows a small floating indicator while a meeting records, with Copy new and
Copy all for the transcript so far.
- Keeps meeting recordings when the display sleeps: Hex finishes and
transcribes the audio it captured instead of losing it.
- Adds a guided first run with a short practice dictation.
- Fixes a slow first dictation on Windows after installing or updating.
- Fixes Linux and Windows CI builds.
This macOS release uses build 20301.
Hex 2.3.02026-10-19
Hex 2.3.0
- Adds Windows (beta) for x64 and ARM64. Hold Ctrl+Space to dictate into any
app. Hex lives in the system tray, starts at sign-in, and updates itself from
signed releases. On x64 it uses the GPU through Vulkan; ARM64 runs on the CPU.
- Streams every dictation on macOS, Linux, and Windows: Hex transcribes while
you speak, so releasing the shortcut finishes in about a tenth of a second,
however long you talked. A 2-minute dictation finishes in 0.1 s instead of
0.9 s on an RTX 5070 Ti and 2.8 s on a MacBook Air M4. If streaming fails,
Hex transcribes the whole recording as before, so nothing is lost.
- Switches English dictation to Parakeet Unified English, the streaming model,
once, if it's already downloaded. Your previous model stays downloaded and
one click away in Settings › Local transcription. If Parakeet Unified isn't
downloaded, or you dictate in another language, Hex keeps your model.
- Marks the model that streams with a Streams tag in the model picker, keeps
its Recommended tag after you choose it, and shows the active model with a
dot instead of a badge.
- Removes Transcribe during pauses: streaming replaces it.
- Fixes dictation pasting nothing into web pages in Chrome 155 and 156, and in
other Chromium browsers such as Brave, Arc, and Edge.
- Stops Meetings from labeling the other side's speech as You when you take a
call on speakers, and uses far less CPU while recording a meeting.
- Adds Shortcut mode (Settings › Dictation): Hold to dictate while the shortcut
is down, or Toggle to press once to start and again to finish. Escape cancels.
- Adds Recording indicator (Settings › Behavior): show the dictation pill at the
top or bottom of the screen, or hide it. On macOS, Settings also shows a
Software updates row with Check now.
- Adds Interface language (Settings › Application): English, Chinese, German,
Japanese, Polish, or Spanish, or follow the system.
- Gives Linux the full app: Modes with app rules on X11, History, microphone
choice, Paste last dictation, Double-tap only, Release microphone while idle,
While dictating, the recording indicator on X11 and Wayland, and Voice Action.
- Refreshes the look of every pane: one consistent text size scale, buttons,
and colors, readable secondary text, and sentence-case labels.
- Makes Settings, Modes, Voice Action, and History the same on every platform.
Windows and Linux settings move into settings.json on first launch; the old
file is kept as a backup.
This macOS release uses build 20300. The public TypeScript SDK remains at
0.3.1.
Downloads
Get the right build at hex.kitlangton.com. The files below are identical mirrors.
- Mac (Apple silicon, macOS 15 or later): HEX-2.3.0-arm64.dmg
- Windows 10 or 11: HEX-Setup-2.3.0-x64.exe or HEX-Setup-2.3.0-arm64.exe. The installer isn't code-signed yet, so SmartScreen may warn; choose More info, then Run anyway.
- Linux (x8664 Arch, glibc 2.44 or later): curl -fsSL https://downloads.hex.kitlangton.com/install-linux.sh | sh. HEX-2.3.0-x8664-linux is the executable it installs.
Hex 2.2.0 for macOS2026-10-16
Hex 2.2.0
- Transcribes faster, especially while your Mac is busy: long dictations that
took several seconds now finish in under two.
- Adds Transcribe during pauses (Settings › Behavior, on by default): Hex
transcribes completed speech while you pause, so releasing the shortcut only
has to transcribe the last few seconds.
- Keeps 200 ms of audio after you release the shortcut so the last word is not
cut off, transcribes a short final phrase together with the speech before it,
and avoids stray periods where you paused mid-sentence.
- Skips silent captures instead of pasting phantom text such as "thank you".
- Marks dictated text as transient so clipboard managers do not record it.
- Pastes correctly with keyboard layouts that switch to QWERTY while Command is
held, such as "Dvorak – QWERTY ⌘".
- Fixes a first dictation being dropped when Release microphone while idle is on
and the microphone was slow to open.
- Adds Meetings (beta, off by default; enable it in Settings). Hex offers to
record when a Zoom, Meet, Slack, Teams, or FaceTime call starts, shows a
timer in the menu bar, and transcribes the call locally as it goes without
repeating the other side's speech from your speakers as yours. Apple
Intelligence titles each meeting and writes a short summary on device. The
Meetings pane adds search, rename, delete, and Markdown export, and
hex mcp lets AI agents read your meeting transcripts.
This macOS release uses build 20200. The public TypeScript SDK remains at
0.3.1.