Current
Beta v0.4.1-beta.4 Current Oct 29, 2026
Beta 4: launches cleanly on newer Linux, releases land faster
This beta is about getting Kalam running and getting new builds to you. On Linux distributions that default to OpenSSL 4, such as Ubuntu 26.10, the app now installs and opens normally instead of failing to start. New beta builds also reach the download page as soon as they are published, so the version listed there matches what is actually available.
Fixed
Linux installs launch again on systems with OpenSSL 4
On distributions that default to OpenSSL 4, such as Ubuntu 26.10, the .deb package used to install but Kalam would fail to start. The installer now pulls in the OpenSSL 3 compatibility library automatically, so the app opens normally after install. You no longer need to install anything extra by hand to get Kalam running.
New beta builds reach the download page right away
When a beta is published, the release process can now read the list of files attached to that release. That lets the download page register the build as soon as it is out, instead of lagging behind. The version you see on the download page is the version that is actually available.
Previous
Beta v0.4.1-beta.3 Oct 28, 2026
Your Polish check and Clean take toggles now stay on
A small update that fixes a frustrating settings bug. If you turned on Polish check or Clean take while Kalam was busy checking your license, the switch could quietly flip back off. Now your choices stick the moment you make them.
Fixed
Polish check and Clean take stay on when you enable them
Turning on Polish check or Clean take could sometimes undo itself a second later, leaving the switch off when you expected it on. This happened when a background license check finished at the same moment you were changing settings on the Models page. Kalam now waits its turn instead of refreshing your settings over the top of your change, so both switches stay exactly where you set them.
Beta v0.4.1-beta.1 Oct 28, 2026
Instant cancel, honest mic status, and no more phantom “Thank you.”
This release makes the dictation pill tell you what's actually happening: it says “Starting microphone” only until real audio arrives, and its buttons respond to clicks on macOS again. Cancel now takes effect the moment you press it, and the app no longer adds an invented “Thank you.” to the end of your dictation. History also shows how long your microphone took to start, and the Sense controls are now three compact icons.
New
See microphone startup time in history
History entries now start their latency bar with a Microphone slice, showing the wait between starting dictation and the first audio arriving. Older entries keep the breakdown they already had. The response-latency number on Home still measures only the speech step.
Fixed
The pill only says “Starting microphone” until audio actually arrives
The label now stays up only while no audio has come in. The first buffer, even a quiet one, switches to the listening wave, so you no longer sit through a fixed delay or have to make noise to get it to change. A microphone that stays silent still shows “No sound from the microphone” after about a second, and the mic still closes when you let go. The Linux panel chip uses the same label.
Pill buttons respond to clicks again on macOS
Expand, Fix, Retry, and Cancel on the macOS pill were ignoring clicks because cursor positions were being compared in the wrong coordinate space. Cursor position is now matched correctly, including on setups with mixed display scaling, so the overlay accepts clicks and appears on the right screen.
Sense controls are now three compact status icons
Clean take, Polish check, and Smart routing were three oversized text buttons; they are now three borderless icons that show a small green dot when on and grey out when off, with the name and state in the tooltip. Clicking one opens Models, expands that card, scrolls to it, and flashes an orange outline so you can find the setting. The icons match the ones on the Models cards.
The Settings button in the sidebar is a proper square
It used to stretch to the height of the account button while sizing its width from the gear, so it looked like a tall rectangle. It is now a fixed 48px square that lines up with the account button.
Cancel takes effect the moment you click it
Pill Cancel, Esc, and panel Cancel now stop Sense, polish, insert, voice commands, capture actions, voice edit, and Hey Kalam clipboard placement right away instead of waiting for those steps to finish. Long transcriptions stop between chunks, retry waits wake up on cancel, and the Cancelling state appears immediately. Cancelling while still recording also stops the microphone.
Undo and Scratch That only work as standalone commands
Saying “undo”, “delete that”, or “scratch that” now triggers only when it is the whole utterance, so those words inside a normal sentence are left alone. When it does fire, it correctly removes the final sentence even when it ends with a period, and keeps your line breaks intact.
No more phantom “Thank you.” at the end of dictation
The app no longer sends a second or more of trailing silence to the speech engine, which was causing it to invent a closing “Thank you.” It also filters out trailing “Thank you.” results that land past the end of your audio, last only a fraction of a second, or sit over silence, while keeping a real “Thank you.” that you actually said. Brief key-click blips are ignored too.
Also
Spoken Punctuation commands have been retired
The old commands for saying “period”, “comma”, “new line”, and similar are gone, along with their card in Commands. Modern speech engines punctuate from how you speak, and AI polish formats paragraphs and lists for you. Punctuation now comes from your phrasing instead of from spoken keywords.
Beta v0.3.0-beta.2 Oct 15, 2026
Kalam 0.3.0-beta.2 Released
This update brings new features like retrying transcription from the pill and profile icons in the pill. It also fixes several bugs, including issues with the overview figures, profile list, and pill layout. The app window now opens at a minimum size and the dictation pill has been redesigned. These changes improve the overall user experience and make it easier to use Kalam for dictation and transcription. The update also includes several other improvements and bug fixes, making Kalam more reliable and efficient.
New
Retry Transcription from the Pill
You can now retry transcription from the pill, which runs the same take through speech-to-text and types it when the audio is still in memory. This feature is useful when you need to retry a transcription due to an error. On the expanded idle pill, Transcribe again runs it again with the selected profile and updates the History row.
Profile Icons in the Pill
Each profile's icon now shows in the pill, making it easier to identify and select profiles. The icon or brand logo you set in Dictation appears beside the profile in the expanded pill's list and in the Recording pill.
Fixed
Improved Overview Figures
The overview figures have been improved to recede on paper, making them easier to read. The selected tab is now key ink on paper, and the range tabs sit below the plate. The overview cards also keep their size while stats load, preventing them from resizing unexpectedly.
Fixed Profile List and Pill Layout Issues
The expanded pill's profile list is no longer cut off at the top, and the list now fits the room and scrolls. The Error and Needs-fix pills have been laid out properly, with the headline sitting on a key plate and the buttons lining up under the headline. The pill's mode list also opens again, and the Transcribing and Cancel labels are now readable.
License and Expiry Issues Fixed
A grace license now stays Pro or Max while the grace window is open, even when the billed period has already passed. Expired licenses are now labeled as Expired, and free and expired licenses cannot create or duplicate a dictation profile.
Read Surroundings Issues Fixed
The read surroundings feature has been improved to skip parent Extended and selection on macOS secure fields. The feature also uses tokens to prevent the prompt fence from closing unexpectedly. On Windows, the feature skips password controls and continues up the parent chain.
Other Bug Fixes
Several other bugs have been fixed, including issues with the dictation pill's layout, the success pill's labeling, and the app's behavior on different platforms. These fixes improve the overall stability and reliability of Kalam.
Also
Minimum App Window Size
The app window now opens at a minimum size of 1075x725, making it easier to use on smaller screens. The setup window also uses the same size.
Reworked Dictation Pill
The dictation pill has been reworked to match the brand, with a new design that includes a printed stage, segmented controls, and a live profile menu. The pill also prints in overprint, with a mix of the profile color and near-black.
Improved Printing and Design
The app's printing and design have been improved, with new features like overprint grounds, two-drum registration, and stippled elevation contours. The pill's recording badges, idle dot, and success check also use registration shadows instead of blur.
Other Changes
Several other changes have been made, including improvements to the app's behavior on Linux and Windows, and changes to the way the app handles private dev tags and trial periods. These changes improve the overall user experience and make Kalam more efficient and reliable.
Beta v0.2.0-beta.3 Oct 12, 2026
Kalam 0.2.0-beta.3 Released
This update brings a range of new features, including scheduled updates, flip-dot Overview figures, and improved dictation profiles. Bug fixes include issues with Linux panel applets, update checks, and dictation polish. The app also features a new design, with a risograph-inspired aesthetic and improved accessibility options.
New
Scheduled Updates
Kalam now checks for updates on a schedule, with beta channels checking once a day and stable channels checking every two days. You can also install updates from the About page.
Flip-Dot Overview Figures
Overview figures now use a flip-dot board design, with discs that turn when the figure changes. This feature is available for Time saved, speaking pace, words, and other key metrics.
Improved Dictation Profiles
Dictation profiles have been redesigned to make it easier to tune your settings. Each profile now shows a journey of stages, including You speak, Hear, Read surroundings, and Improve. You can also load live model lists from each provider.
New Language Support
Kalam now supports English, French, Arabic, and Spanish. You can switch languages from the Settings page, and the app will remember your choice.
Redesigned Dictation Profile Screen
The dictation profile screen has been redesigned to make it easier to use. The name is now the page title, and the accent colors sit in a row under it. The pipeline is a vertical route with stage cards for Hear, Read surroundings, and Improve.
Fixed
Linux Panel Applets Restored
Linux panel applets are now working again, after a previous issue caused them to disappear. This fix ensures that the GNOME extension and other DE chips are available.
Update Checks Improved
Update checks have been improved to fix issues with failed checks and incorrect notifications. The app will now log warnings and show error notifications if a check fails.
Dictation Polish Issues Resolved
Issues with dictation polish have been resolved, including problems with token caps and audio ducking. The app will now ask for up to 4096 tokens and type the speech text if the provider stops due to a token cap.
Window and Dock Icon Behavior Fixed
The window and Dock icon behavior has been fixed, so that the Dock icon stays near the clock and the window comes back when you open Kalam from the icon.
Update Button on Windows Fixed
The Update button on Windows has been fixed, so that it no longer offers to update on next start. Instead, the app will update immediately.
Also
New Design and Aesthetic
Kalam has a new design and aesthetic, inspired by risograph printing. The app features a range of new visual elements, including a hoopoe mark, Fraunces titles, and Source Sans 3 fonts.
Improved Accessibility Options
The app now features improved accessibility options, including reduced motion and high contrast mode. These options can be enabled from the Settings page.
Settings and Theme Switcher Changes
The Settings page has been redesigned, with a new square button and a theme switcher in the bottom bar. The plan and version information now open a page, and the engine sits beside the dictation mode.
Beta v0.2.0-beta.2 Oct 8, 2026
Kalam 0.2.0-beta.2 Released
This update brings new features and bug fixes to Kalam, including the ability to remove Kalam from your Mac and use an xAI key for dictation. The dictation pill has also been improved, and issues with microphone detection have been resolved. Additionally, permissions on macOS have been updated to survive updates and ensure a smoother experience.
New
Remove Kalam from Mac
You can now remove Kalam from your Mac through the About section in Settings. This will also give you the option to quit Kalam, remove it from your login items, and move it to the Trash. Your settings, history, and downloaded speech models will remain unless you choose to remove them as well.
xAI Speech Engine
Kalam now supports the use of an xAI key for dictation. To use this feature, simply hold the dictation keys as you normally would, and the recording will be sent when you let go. The xAI key will remain in the app for future use.
Fixed
Dictation Pill Improvements
The dictation pill has been improved to provide better feedback when the microphone is not sending sound. If you hold to dictate and the microphone is missing, muted, or silent, the pill will let you know while you are still holding, and again when you let go. Additionally, the pill will no longer blink off during a hold, and starting dictation on Mac will no longer quit the app.
Microphone Detection
Kalam now correctly detects when the microphone is not sending sound, and will let you know if you are in a quiet room or if the microphone is muted or missing. The app will also start dictation while the keys are being held, rather than waiting for them to be released.
macOS Permissions
Permissions on macOS have been updated to survive updates, ensuring that Accessibility and Input Monitoring stay granted after an update. If a grant was saved for an older copy of Kalam, the app will clear it once at startup when it is not actually granted, and Access and Settings will offer Repair instead of asking you to restart.
Also
Local Mac Flicker Probe
A new feature has been added to help diagnose issues with the dictation pill on Mac. By setting the KALAM_OVERLAY_TRACE environment variable to 1, you can record overlay window calls and capture the pill and line blank frames. This feature is not part of a normal install and is intended for debugging purposes only.
Beta v0.1.0-beta.36 Sep 15, 2026
Kalam 0.1.0-beta.36 Released
This update adds new speech providers, including Google AI Studio, Deepgram, AssemblyAI, and ElevenLabs, and fixes several issues with dictation and keyboard input. It also includes changes to the settings and speech provider marks.
New
More Speech Providers
We've added several new speech providers to Kalam, including Google AI Studio, Deepgram, AssemblyAI, and ElevenLabs. These providers offer high-quality speech recognition and can be used with an API key. This means you'll have more options for dictation and can choose the provider that works best for you.
Fixed
Dictation Issues Resolved
We've fixed several issues with dictation, including a problem where the dictation pill would leave behind a 'ghost' after use. We've also resolved issues with keyboard input, including a problem where the Control key wasn't working correctly in some fields. These fixes should make dictation and keyboard input more reliable and easier to use.
AssemblyAI Long Holds Fixed
We've fixed an issue with AssemblyAI where long holds would cause the take to fail. Now, sync listen clips will be uploaded to the catalog after a certain time limit, rather than failing the whole take. This should make it easier to use AssemblyAI for dictation.
Gemini Speech Fallback Added
We've added a fallback for Gemini speech recognition. If Interactions returns completed with no transcript fields, Kalam will now fall back to multimodal generateContent on Flash-Lite with the same AI Studio key. This should improve the reliability of Gemini speech recognition.
Also
Custom OpenAI-Compatible Endpoint Removed
We've removed the custom OpenAI-compatible endpoint from the settings. This means you'll no longer be able to enter a free-form OpenAI-compatible URL. However, OpenRouter remains available for OpenAI-compatible polish. Existing custom OpenAI profile LLMs will be remapped to inherit on config v26.
Speech Provider Marks Updated
We've updated the speech provider marks for Deepgram and ElevenLabs to use Simple Icons, and AssemblyAI now uses the Lobe Icons mark. You can find more information about the icons and their attribution in the public/logo/providers/ATTRIBUTION.txt file.
Beta v0.1.0-beta.32 Sep 3, 2026
Kalam 0.1.0-beta.32 Released
This update brings new features like a bundled STT/LLM catalog, improved history latency bars, and a Linux panel applet. It also fixes several bugs, including issues with local providers, context, and the Linux panel. The update also includes changes to the OpenAI polish default and the History latency bar display.
New
Bundled STT/LLM Catalog
Kalam now comes with a bundled catalog of recommended models for Groq, OpenAI, Anthropic, Gemini, OpenRouter, and custom endpoints. This makes it easier to get started with dictation and find the best models for your needs. The catalog is stored in a file called model_support.json and is used to populate the dictation provider list.
STT Dispatcher
The STT dispatcher now packs adjacent pauses towards 30 seconds for Groq Whisper and local Whisper, and sends one request for OpenAI gpt-transcribe and gpt-4o-mini-transcribe. This improves the efficiency of dictation and reduces the number of requests made to the STT providers. Dictionary terms are also sent as keywords plus prompt to gpt-transcribe.
History End-to-End Latency Bar
The history latency bar now shows where the wait went for each dictation, including Engine, Speech, Context, Polish, and Insert. This allows you to compare STT and LLM providers with real numbers and see where the bottlenecks are in your dictation workflow. Older rows still show a single Speech slice from the saved STT time.
History Context Section
The history context section now expands sources that had text, and shows a single line for sources that had no text. New takes store the excerpt polish used, and older takes still list names without text. This makes it easier to see the context for each dictation and understand how the models are being used.
Linux Dictation Pill as a Panel Applet
The dictation pill is now a panel applet on Linux, showing the live recording waveform, mode name, and timer, with a click-to-stop and a menu for modes, dictation on/off, and Open Kalam. This makes it easier to access and control dictation from the Linux panel.
Applet Page for Linux
The applet page for Linux now shows a mock panel strip, a 2.5× magnifier, state tabs, and a Live toggle that mirrors what the daemon is doing right now. This allows you to customize and configure the applet to your needs.
Three Waveforms in the Panel Chip
The Linux panel chip now draws three waveforms: Ribbon, Hairline, and Bars. The Applet page and browser lab pickers label them accordingly. This gives you more options for customizing the appearance of the dictation pill.
Copy Last Transcription from the Linux Panel Menu
The Linux panel menu now has a Copy last transcription option, which copies the last history take through the daemon. This makes it easier to access and use your dictations.
Linux Panel Transcribing Matches the Overlay Retry Chrome
The Linux panel transcribing now matches the overlay retry chrome, showing elapsed seconds, and turning gold with “Still working…” or red with “Trying again… (N)”. This provides a consistent and intuitive interface for dictation.
Fixed
Local Providers No Longer Ask to Install Unavailable Engines
Local providers no longer ask you to install an engine that cannot run on your system. Instead, they explain why the engine is unavailable. This prevents confusion and frustration when trying to use dictation.
Linux Context No Longer Claims Focused-Field Text
The Linux context no longer claims focused-field text, treating Focused/Extended as window title only. This fixes an issue where the context was incorrectly including text from the focused field.
COSMIC Click-Away No Longer Claims Inserted
The COSMIC click-away no longer claims Inserted, only typing if the original field activated IME. Otherwise, the transcript is copied and the chip shows Copied — paste. This fixes an issue where the dictation was being inserted incorrectly.
Linux Auto-Activate Picker No Longer Hides Cursor and Claude
The Linux auto-activate picker no longer hides Cursor and Claude, searching name + WM_CLASS + Exec stem, and keeping both com.anthropic.claude and claude-desktop as identity keys. This fixes an issue where certain applications were not being detected correctly.
Linux Wayland Open-Windows Picker No Longer Empty
The Linux Wayland open-windows picker is no longer empty, listing compositor toplevels from zcosmic_toplevel_info and ext_foreign_toplevel_list when advertised. This fixes an issue where the picker was not showing any windows on Wayland.
Linux Auto-Activate App Search No Longer Lists Electron Threads
The Linux auto-activate app search no longer lists Electron threads, listing application-drawer .desktop entries instead. This fixes an issue where unnecessary threads were being shown in the search results.
COSMIC Panel Quit No Longer Relaunches Kalam
The COSMIC panel Quit no longer relaunches Kalam, only allowing Open Kalam to activate the daemon. This fixes an issue where the daemon was being restarted unnecessarily.
Linux Applet Page Chip Follows the Dictation Profile Color
The Linux applet page chip now follows the dictation profile color, converting OKLCH to sRGB on the bus. This fixes an issue where the chip was not matching the profile color.
Linux Panel Idle Dash and Recording Waveform Follow the Dictation Profile
The Linux panel idle dash and recording waveform now follow the dictation profile, converting OKLCH to sRGB. This fixes an issue where the dash and waveform were not matching the profile color.
COSMIC Panel Chip Hover No Longer Sits Far from the Panel Edge
The COSMIC panel chip hover no longer sits far from the panel edge, filling the panel slot with a 1 px inset. This fixes an issue where the hover was not aligning correctly with the panel.
COSMIC Transcribing No Longer Shows Three Static Dots
The COSMIC transcribing no longer shows three static dots, drawing a rotating circular notch instead. This fixes an issue where the transcribing indicator was not animating correctly.
Linux Panel Transcribing Spinner No Longer Clipped
The Linux panel transcribing spinner is no longer clipped, with a path radius that is 90% of the old size. This fixes an issue where the spinner was being cut off.
COSMIC Panel Left Stroke Crumbs on the Waveform and Spinner
The COSMIC panel no longer leaves stroke crumbs on the waveform and spinner, dirtying the whole canvas each tick. This fixes an issue where the waveform and spinner were not being redrawn correctly.
Also
OpenAI Polish Default is Now GPT-5.6 Luna
The OpenAI polish default is now GPT-5.6 Luna, which provides faster and more accurate results. The saved gpt-4.1-mini remaps on config v25, and polish sends reasoning_effort: none to stay fast.
Kalam Max is No Longer for Sale
Kalam Max is no longer for sale, and the Settings hide the Kalam Cloud card. Onboarding no longer treats Max as a ready engine, and dictation pickers do not re-inject leftover kalam_cloud.
History Latency Bar is Easier to Read
The history latency bar is now easier to read, with a heading that says End-to-end latency, and Speech and Polish using chip colors. Waits of 1s or more show as seconds, and hover shows the exact millisecond count.
History Context is Now its Own Section
The history context is now its own section, with sources that had text expanding, and empty attempts showing as a single line. Features used is removed, and those facts stay on the chips.
Linux Panel Copied is Not Inserted
The Linux panel Copied is not Inserted, showing a green check + Copied — paste instead. This provides a clear indication of when the dictation has been copied to the clipboard.
Linux Panel Inserted is a Green Check + Word Count
The Linux panel Inserted is now a green check + word count, showing the number of words inserted. This provides a clear indication of when the dictation has been successfully inserted.
Linux Panel Error is a Red Warning + Error
The Linux panel Error is now a red warning + Error, providing a clear indication of when an error has occurred.
Linux Panel Idle is Dash + Profile Name
The Linux panel idle is now dash + profile name, showing a short accent rule plus the mode name. This provides a clear indication of when the dictation is idle.
Linux Panel Waveform is Sharper and 60 fps
The Linux panel waveform is now sharper and 60 fps, using 96 samples and a wave box that is icon × 1.6 tall. Preview redraws every 16 ms, and phase step and RMS smoothing are scaled to match the old motion speed.
Linux Floating Overlay HUD Archived
The Linux floating overlay HUD is now archived, and the X11 override-redirect WebKitGTK pill is no longer built or shown on Linux. This simplifies the codebase and reduces maintenance.
Beta v0.1.0-beta.31 Sep 1, 2026
Dictation Pause Fix
This update resolves an issue with dictation pause and Quiet mode. It ensures that your local changes are saved correctly and won't be overwritten. This fix improves the overall dictation experience.
- Quiet mode now turns off correctly when pausing dictation
- Local changes are saved without being overwritten
Beta v0.1.0-beta.27 Aug 31, 2026
Kalam Update
This update introduces a system tray menu, improved dictation features, and numerous bug fixes to improve overall performance. The dictation pill can now be resized, and several issues with the Linux version have been resolved. Additionally, the update includes improvements to the user interface and accessibility features.
- - System tray menu added
- - Dictation functionality improved
- - Bug fixes for enhanced performance
Beta v0.1.0-beta.26 Aug 17, 2026
Improved Model Interactions
This update addresses issues with language hints and cloud LLM catalog IDs. It ensures that the app works correctly with various models, including OpenAI and Anthropic. The update also remaps saved profile IDs to their new defaults, preventing potential errors.
- Fixed language hint issues with OpenAI models
- Updated cloud LLM catalog IDs to prevent shutdowns
Beta v0.1.0-beta.25 Aug 14, 2026
Kalam Update Released
The latest update to Kalam includes new voice actions, improved local speech-to-text models, and a revamped settings page. Users can now access new features like voice actions, which allow for hands-free control, and local STT models, which provide more accurate transcription. The settings page has also been redesigned for easier navigation and management of cloud and local providers. Additionally, various bugs and issues have been fixed to improve overall performance and user experience.
- New voice actions for hands-free control
- Improved local speech-to-text models
- Redesigned settings page for easier navigation
Download free 