Every published release of Local Waifu, newest first, in plain language. For the full technical notes and downloads, head to the releases page on GitHub.

v1.8.0 November 4, 2026 macOS Windows

Her 3D body, on your desktop

She has a 3D body now. Give her a VRM model (from VRoid Studio or VRoid Hub) or an MMD model (.pmx or .zip) and she blinks, breathes, follows your cursor with her eyes, moves her lips when she speaks, shows 16 feelings and makes gestures on her own while she talks. Ask her in chat to wave, sit down, dance, lie down or make an angry face. Settings → 3D Avatar has a live preview beside every option, backgrounds, poses with a pose editor, animations, and a guide to getting your own character. Six sample avatars from VRoid Hub come built in.
You can import your own animations. A .vrma file works as a gesture or as a full-body animation such as a dance or a walk, and your dances join "dance for me".
A desktop pet mode puts only her figure on your desktop (Option+Cmd+5 on a Mac, Ctrl+Shift+5 on Windows). She walks, sits on the edge of the Dock or the taskbar and dances, and a right-click menu gives you poses, animations, size and "Stay on top". Photo mode and selfies from her 3D model are new too.
She goes quiet by herself during meetings, full-screen apps and when the screen is locked. Anything she wanted to say is held and comes later.
Memory map, screen awareness, routines, group chats, voice messages, hands-free "hey <name>", bedtime stories, songs, localwaifu:// shortcuts and the visual novel mode are now on both macOS and Windows. A short in-app guide covers the basics.
The cloud is a fallback only if you turn it on. When the local engine is down, or she gets a photo no local model can see, the message used to go to any cloud key you had saved. Now that needs a switch in Settings, next to Offline Mode, and it is off by default.
macOS 14 Sonoma or newer is required. The bundled AI engines already needed it, so on macOS 13 the app installed but local chat and pictures could not start.
Pressing Stop now really stops her. The local model used to keep writing the rest of her reply in the background, so your next message waited behind it. Replies also start sooner on slower computers, and 8 GB Macs begin on a smaller model that no longer makes the whole machine swap.
Recovery got safer. A new "Start fresh, keep a copy" option on the recovery screen moves your whole data folder aside untouched, and the local engine restarts by itself if it crashes. Opening the app a second time no longer touches the one that is running. Reset all data now works.
Your licence key can no longer be read or edited from the app window, and the logs no longer contain her memories or the sentences she speaks.
Back from the desktop pet on a Mac, the window frame is restored properly: no more empty strip above the app with the three window buttons floating on it.
On Windows, AMD graphics cards are read for the model recommendation, and Settings → Hardware shows the adapter, its memory and whether ROCm or Vulkan acceleration is ready. The bundled AI engine is now Ollama 0.40.13.

v1.7.6 October 29, 2026 macOS Windows

Gemma 4 downloads work again, and the engine can update itself

Downloading Gemma 4 failed with "requires a newer version of Ollama". The Ollama model registry stopped accepting the engine version we shipped (0.30.8) for every Gemma 4 model, so the recommended model for 8-16 GB computers could not be installed at all. The app now ships Ollama 0.40.12, and Gemma 4 installs normally.
The engine can now be updated without waiting for a new app release. The app checks a signed list on localwaifu.com once a day. A newer engine is downloaded only from Ollama's official GitHub releases, checked against its signature and checksum, tested, and then used from the next start. If a new engine ever fails to start, the app goes back to the bundled one immediately, in the same launch.
When the engine is too old for a model, the app now says so plainly. If the engine is one it manages, it updates the engine and tries the download again. If the Ollama answering is one you run yourself, or one on another computer, the message says which, and what to do. Settings → Advanced shows which engine version is running, with a button to check for an update.
If Gemma 4 still can't be downloaded during first setup, the app offers Qwen 3.5 4B instead, so setup can finish.
On Windows, AMD graphics cards get GPU acceleration for chat. Before, the AMD download was missing the engine itself, so AMD computers always ran chat on the processor. Every engine file is now checked for Ollama's digital signature before it runs.

v1.7.5 October 20, 2026 macOS Windows

Edit her prompt, edit your messages, paste images

Settings → Developer lets you read and rewrite every block of the instructions she gets before each reply: her identity line, how she writes, the voice-call rules, her memory and photo tools. Each block is marked Safe, Caution or Dangerous, and the dangerous ones (the parts that teach her how to save memories, send photos and switch moods) ask you to confirm before saving. A preview shows the prompt as she would get it right now. If you never touch it, she gets exactly the same instructions as in 1.7.4.
The Developer tab also has your own blocks for standing requests like "always call me Captain", the generation settings (temperature and friends), presets you can save and switch between, and export and import of the whole setup. A "Restore everything to defaults" button at the top puts it all back in one click; your saved presets are kept.
You can edit your own messages. Change one and she answers it again: the messages after it are removed, along with anything she remembered only because of them.
Paste a screenshot with Cmd+V / Ctrl+V or drag an image anywhere onto the window to attach it. A photo can be sent on its own, without text, and she reacts to what she sees. Images over 5 MB or in formats the models can't read, such as HEIC, are resized and converted instead of refused, so a full-screen screenshot from a high-resolution display goes through.
An emoji button in the message box. Picks go where your cursor is, and the picker stays open so you can add a few in a row.
The setup screen lets you choose your local model. Besides the one recommended for your computer, it now lists every local model the app supports with its download size and whether it can see photos; models your memory can't run are greyed out.
Download sizes are correct. Several models showed the wrong size before you downloaded them; the worst was Llama 4 Scout, listed as 11 GB when it is a 67 GB download that needs 96 GB of memory. Every size now comes from one list, checked against Ollama's own registry.
Typing in Japanese, Chinese or Korean no longer sends a half-written word when you press Enter to pick a character.
The view-mode shortcuts (Cmd/Ctrl+Shift+1, 2, 3) work on every keyboard layout, including German, where Shift+2 and Shift+3 type different characters.
OpenRouter's "Auto Router" no longer shows up as an image model. It picks a model per request and could pick one that doesn't draw.
On Windows, a large graphics card now counts when Settings and setup judge which models fit, the same way the recommendation already did: a 16 GB PC with a 24 GB card no longer sees models it runs fine marked "needs 32 GB". The voice and cloud-key settings also stopped saying "this Mac", and the call screen shows Ctrl+Shift+C instead of ⌘⇧C.

v1.7.4 October 10, 2026 macOS Windows

Back to your latest message, and safer updates

The chat now opens at your latest message. It used to open at the top of the loaded history at launch, when coming back from the Memory Book or the Gallery, and after switching window modes, and switching to a longer conversation dropped you in the middle of it. It also stays at the bottom while a photo she sent finishes loading, and still leaves you alone while you scroll up to read.
Search takes you to the message. Cmd+F / Ctrl+F finds messages across all of a character's chats; clicking a result from another chat, or one older than the latest 50 messages, now opens the right chat, loads enough history and scrolls straight to it.
A failed "Try again" no longer hides her reply: her original answer stays on screen and the error shows as a notification.
After Reset Database, creating a character asks your name first, and the chat shows a clear "create a character" screen. The character creator, photo cropper and feedback windows scroll on small windows.
Licenses have no device limit. Settings → License no longer has a "Deactivate this device" button or a transfer guide: on a new computer, install Local Waifu and redeem the same key.
Only one in-app update can install at a time, even if Install is clicked from both the notification and Settings. Download progress reporting is smoother, and the update controls in Settings are translated. On Windows, the 1.7.4 updater checks downloaded installers against the release signature before they can run and deletes a file that does not match. Direct-edition updates require a license; store editions continue updating through their store.
On Windows, Offline Mode also blocks cloud-provider key tests and live model lists. Setup checks a cloud key before saving it, Image Models lists cloud image providers again, and dropping a non-image file into chat now explains the problem.
Windows translations now cover the adaptation panel, error screen, knowledge browser and missing Image Models text in all seven app languages. Reset Database on Windows also removes generated photos and portraits; this is a destructive reset, not part of installing an update.

v1.7.3 September 23, 2026 macOS Windows

Edit her replies, plus a proper itch.io edition

Every reply of hers now has a small pencil under it. Click it and the bubble becomes a text box holding her exact words: Enter saves, Shift+Enter starts a new line, Escape cancels, and Cancel leaves what you typed exactly where it was. It fixes a typo, a wrong name, the wrong language, or a line that should never have been in the transcript, without throwing the whole turn away and re-rolling it.
The correction reaches her. Your transcript is replayed to her as her own past words on every later turn, so from the next message on she reads the corrected version rather than the original.
An edited message is marked "edited" beside its timestamp, permanently and in all seven languages, so a rewritten transcript never passes as her verbatim words. The save notice also offers Undo, which puts the previous text back.
Editing does not rewrite her memory. A memory she saved from that message stays, and the notice after saving says so and offers the Memory Book, where you can edit or delete it yourself. An edit also never awards relationship experience and never sends anything to a model.
The itch.io download is now a dedicated unlocked edition. Buying it on itch.io gives you the complete app without a separate trial, licence key, paywall, checkout screen, or subscription. Updates for that edition continue through itch.io.
The direct edition checks for updates when it starts again. Nothing checked for a new version at launch since 1.7.0, so the only way to learn one existed was Settings → General → "Check for Updates". The app now checks once, a few seconds after launch, and shows a notification when a newer version exists, with a single Install action that downloads the signed build and restarts into it. Nothing is downloaded or installed until you click it.
Settings → General → Updates has an Auto Update switch. Off means no launch-time check at all, and its description now says what it actually does: check for new versions when the app starts. Updates remain a lifetime-customer feature, and the Mac App Store and itch.io editions keep receiving their updates through their store.

v1.7.2 September 8, 2026 macOS Windows

Your own image models, and a steadier local companion

Bring your own full SDXL .safetensors checkpoints and LoRAs from Settings → Models → Image Models. They are copied into a private local library and never uploaded; each checkpoint can use up to three LoRAs with their own weights.
Photo Director now keeps the right companion in the frame. It uses the tested local identity path, notices when her saved portrait changes, and discards a late result if you switch character or portrait while it is rendering.
On Windows, local chat now protects itself when a long prompt would overflow a small context window. It keeps the current message and the important parts of the conversation instead of returning an empty or one-word reply.
Cloud replies are more resilient, and cloud voice on Windows now handles voice IDs containing spaces, slashes, or question marks correctly.
Memory lookup is faster on Mac again: the search index is used instead of silently falling back to a full scan.
Several smaller Image Models fixes are included too: safer imports and deletion, clearer progress and recovery messages, reliable LoRA controls on Mac, imported models in Telegram's /imgmodel picker, and an API-key field that no longer claims a saved key is missing.

v1.7.1 August 25, 2026 macOS

A repair release for anyone who updated inside the app

If image generation stopped working after an update, and said the local image engine was not bundled in this build, this release fixes it. Updating to 1.7.1 repairs the installation on its own.
The same problem could leave local chat replies empty, because the part that actually runs your local model was missing too. It comes back with this update as well.
Settings showed the local image engine as not installed and offered no way to download a model, because that download button is hidden whenever the engine is missing. It is back once the engine is.
The cause was in packaging rather than in the app. The update package was assembled before the local engines were placed inside it, so it shipped at 70 MB where the installer is 245 MB. Installing from the .dmg was never affected, and neither were your characters, chats, licence or downloaded models.
The build now assembles the update from the finished app, and stops with an error if any engine is missing from it.
Windows is unaffected and unchanged. It carries the same version number because both platforms share a single update manifest.

v1.7.0 August 24, 2026 macOS Windows

More choice, steadier local runtime

Cloud chat, image, speech, and transcription choices are expanded, while local models remain the default. Every cloud service is optional and selected by you in Settings.
Quitting the Mac app now cleans up local model, voice, and image-render processes instead of leaving them behind to hold memory or ports.
Scheduled check-ins wait for a natural pause instead of landing in the middle of your reply.
Background memory work uses the same context size you chose for chat, so it does not reload the local model at a smaller window right after startup.
Windows gets the same provider safety work plus a safe Custom context setting from 512 to 131,072 tokens.

v1.6.3 August 14, 2026 macOS Windows

She remembers again, and setup feels right

Her memory model now downloads when the app starts, so she can recall the things you have shared instead of answering as if you had never spoken. Memory search also stays quick while it is still downloading or when you are offline.
Switching characters while a screen is loading no longer mixes up their lore, knowledge, voice, or edits. A failed chat-list load also cannot open a stray empty conversation.
First setup is smoother. The voice setup step opens when it is needed, the three-step character creator counts correctly, and an uploaded portrait no longer gets replaced by a preset.
The 18+ and AI-companion confirmation is now translated in every supported language. Several screens that were stuck in English are translated too.
The Windows edition catches up with the same release. Boot, offline mode, image generation, and the memory path now share the fixes from the Mac build.

v1.6.2 August 5, 2026 macOS

A thorough pass on stability, privacy, and data safety

Private cloud access tokens are fully hidden in error messages, including the kind that used to leave part of a bearer token visible.
Backing up your data, resetting the app, and starting the Telegram bridge no longer hold up the rest of the app while they wait on the disk or network.
A rare upgrade problem with duplicate relationship data is repaired before it can stop the app from opening. A damaged lorebook entry also stays safely out of your chat instead of becoming active everywhere.
Her adaptive photo learning now keeps conversations separate, so a photo request in one chat cannot be paired with a reply from another.
The natural voice no longer stays stuck as unavailable after it finishes installing. Several smaller recovery, export, and accessibility fixes are included too.

v1.6.1 August 4, 2026 macOS

Safer backups, steadier calls, and clearer first steps

Backups are now crash-safe, and Reset all data can restore everything if it is interrupted instead of leaving part of your data deleted.
A voice call cannot leave image generation or background thinking blocked after it ends. The app also cleans up stuck voice and local-model processes more reliably.
Regenerating a reply always targets the right message, even if another reply arrives at almost the same time.
Before you begin, the Welcome screen now asks for a simple 18+ and AI-companion confirmation with links to the Terms and Privacy Policy. Settings also says plainly that her text, voice, and pictures are AI-generated.
Several loading and chat races are gone, so an older screen response cannot overwrite the conversation you just selected.

v1.6.0 August 2, 2026 macOS

Voice calls, faster image generation, and better text replies

She can now hold a real spoken conversation. Talk to her out loud and she talks back, in a voice that matches her personality. Choose from natural voices, cloned voices, and curated voice styles per character, all in the new Voice screen in Settings. Cloud speech-to-text is available if you want it.
Image generation got faster and more reliable. Her selfies and pictures render with better quality, and she stopped sending headless photos — the image instructions now describe what should be in the shot instead of what to avoid.
Text generation improved across all local models. Her replies come out cleaner and warmer, especially on the bigger models. And she uses what she remembers about you more often, even on smaller Macs.
Dozens of bugs fixed. License errors now explain themselves in your language instead of an English code. Pace, pitch and expressiveness controls actually work. The broken customer portal button is gone. Voice stays set after restart. And a lot more under the hood.
Windows stays at version 1.5.1. The Windows edition does not have voice mode yet — it will catch up in a future release.

v1.5.1 July 19, 2026 macOS Windows

Much faster local chat, and two new looks

Local replies are dramatically faster. Answers that used to crawl, sometimes past a minute, come back in seconds now. She had been quietly overthinking every message out of sight, and that is fixed.
Your message always goes first. When she is tidying up her memory in the background, your next message no longer waits behind it.
The first reply after you open the app is quick now. She warms up the moment she launches instead of pausing on your first message, and on Macs with plenty of memory she stays ready for hours.
The bigger local models sound better, running with their own recommended settings, so the writing comes out cleaner and warmer.
Two new local image styles to pick from: a true anime look and a photorealistic one, alongside the ones already there.
Clear out image models you no longer use, right from Settings, and it tells you how much space you get back.

v1.5.0 July 15, 2026 macOS Windows

She can draw anything, and selfies look like her

Ask her for any picture, a dragon, a landscape, fan art, and she draws it. Not just selfies anymore.
Her selfies finally look like her. The portrait you picked goes along as a reference, so her face, hair, and eye colour carry over instead of a stranger's. She also stopped growing a third arm when she's holding something.
When she says a photo is coming, it arrives. She understands now that it only reaches you once she actually attaches it.
Hide the chat list on the left, or her profile on the right, from the top bar, and give the conversation the whole window. She remembers how you like it.
The progress bar tells the truth: pictures she draws on your own machine show real step-by-step progress, and cloud ones sweep along instead of freezing half full.
Setup fits smaller windows, and shrinking the main window no longer hides the message box.
Buy once, update forever. The yearly renewal is gone: your license now includes every future update and all support, with no time limit. If you already own her, this is already yours.

v1.4.3 July 2, 2026 macOS Windows

Faster replies, and a thinking dial

Local replies come back noticeably faster now, with no loss in how thoughtful she is.
A new Deeper thinking control, in Settings under Performance: leave it off for quick, direct replies, or turn it on when you want her to slow down and think things through.
Smoothed out a couple of rough edges, so the model settings screen always opens and her replies no longer get cut short.

v1.4.0 July 1, 2026 macOS Windows

Five new languages, and she can see

The app now speaks seven languages: English, Polish, Simplified Chinese, Korean, Japanese, German, and Spanish.
On first launch it picks up your system language automatically. Change it anytime in Settings or during setup, and she replies in the language you choose.
Share a photo in chat and she can see and react to it.
She can send you a selfie that matches how you pictured her.
Browse and install any Hugging Face chat model from inside the app.
More cloud model options: OpenRouter, Mistral, Groq, Google, and custom endpoints.

v1.3.1 June 16, 2026 macOS Windows

Windows joins the party

Local Waifu now runs on Windows 10 and 11, published right alongside the Mac version.
One update channel for both platforms, so Mac and PC always stay in step.
New self-check in Settings → Hardware tells you whether she's running on your graphics card or the processor, with a one-tap retry if she isn't.
Models settings on Windows now match the Mac: see what's installed and pick models from your connected accounts.
On Windows, onboarding always shows the portrait you picked.

v1.3.0 June 15, 2026 macOS

Sharper memory, more personal voice

Upgraded, multilingual memory so she recalls more, in more languages. Existing memories re-index themselves quietly in the background.
The personality you set during setup now truly shapes how she talks.
She reaches out on her own with real messages, not just a notification.
Your conversation now survives a restart.
A 'try again' button regenerates her last reply in place.
Faster replies on capable Macs, plus a security and dependency refresh.
Updates are manual now: check for them from Settings whenever you like.

v0.1.21 May 25, 2026 macOS

Mac menu and safety nets

Native macOS menu bar, with Check for Updates right where you'd expect it.
A manual download fallback for when auto-update can't reach the release page.
Dismiss any notification with a click.

v0.1.20 May 25, 2026 macOS

Payments go live

Real checkout: $25 once, lifetime license, key delivered by email.
Clearer Buy buttons and honest, plain-language refund wording.

v0.1.19 May 21, 2026 macOS

She gets an inner life

Named moods she settles into for a few turns: tender, playful, protective, sleepy, and more.
She develops her own likes, dislikes, little quirks, and quiet worries over time.
A weekly self-reflection that gently shapes who she's becoming with you.

v0.1.18 May 21, 2026 macOS

Deeper memory and a Growth view

Relationship stages, from stranger to bonded partner, that change how she talks to you.
Memory that scales with how long you've known each other.
A new Growth dashboard, plus a Memory Book that shows where she learned each thing.

v0.1.16 – 0.1.17 May 2026 macOS

Stability and privacy hardening

A run of internal quality passes: stronger privacy safeguards, crash-safety, and tighter resource limits. No new features, just a sturdier app.

One signed release for both platforms. Both editions check for a new version a few seconds after launch and wait for you to click Install; on Windows you can also press Check for updates in Settings.

Download Local Waifu