What’s new

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Your private files, sealed. Your custom prompt, instructions, memories and chats can now be sealed on disk. A sealed file copied off your PC can’t be read anywhere else, not even by YUKI on another PC. A second switch keeps an item for this PC’s own models only: cloud brains can’t read it, and the brain picker greys them out in those chats. Both are off by default, so your files stay plain and easy to point your own tools at.

A proper goodbye. Uninstalling YUKI now asks what to keep: your chats and custom prompt, and your models. Both are ticked by default. YUKI never deletes anything it didn’t make, and a reinstall finds what you kept.

Bigger brains on more PCs. The newest local runtime is in the catalogue, so very large models can run from your disk on PCs with plenty of RAM. Each model also tunes itself to your PC after its first load.

Film edits you can change. Every film keeps an edit list that you and Yuki can change, down to recutting or rerolling a single shot.

The temperature slider only where it works. Cloud models that don’t take a temperature no longer show the slider.

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

dex inference as your brain, with all of YUKI. With dex inference picked, Yuki can now use her tools: they run on your own PC, never on the host, so she can still open apps, read your files when you ask and work in your browser. Pictures can come from dex inference too. If no host can take tools right now, she tells you and offers your other brains instead of failing quietly.

The thinking slider on dex inference. It’s switched on now. Your plan decides how far it goes; a step above your plan uses one of the day’s Deep thinks.

Every brain on the runtime that suits it. Each model’s runtime and what it needs now comes from the catalogue, for the models we list and for the ones Yuki installs herself. A big model only splits between your graphics card and RAM when it truly doesn’t fit, and when there isn’t room, YUKI says what it needs instead of running slowly.

Finds your tools again. If Windows started YUKI with an out-of-date list of program folders, she now looks up the current one herself, so your other AI apps and tools show up as they should.

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Crash reports, so we can fix what breaks. If YUKI crashes, it sends a short report: what broke, where in YUKI’s code, and the YUKI and Windows versions. Personal details are scrubbed out on your PC first, and it never carries your chats, files, prompts or keys. Turn it off any time in Settings (“Send crash reports”). The Privacy Policy says exactly what’s sent.

Tighter update safety. The saved model catalogue is checked against our signing key every time it’s used, not only when it’s downloaded.

Fixes.

• A remade film clip gets its own thumbnail instead of the old take’s.
• A picture written into an example (inside a code block) no longer hides the real one later in the chat.
• The plans link in a dex inference limit message is clickable.
• The thinking slider scrolls with a swipe on touch screens.
• Notifications no longer write their text into YUKI’s log.

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Pictures and scenes show up the moment they’re done. In a batch, each picture appears in the chat as soon as it finishes, and so does each scene of a film, instead of all at once at the end. When the job is done, nothing shows up twice.

A thinking slider. For brains that think, pick how hard they think: Quick, Balanced, Deep or Max. On dex inference, the slider arrives once it’s switched on there; then your plan decides how far it goes, and higher plans unlock more.

Signed updates. From this build on, YUKI checks every update and the model catalogue against our signing key before it installs anything, and refuses what doesn’t match.

Faster first words in long chats. The start of every request is now the same from turn to turn, so the brain reuses its work instead of reading the whole chat again.

Fixes. Some commands were registered twice at startup; now each is registered once.

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Free up room for brains. Before a brain loads, YUKI first frees what’s hers: an idle picture engine, other brains, voice models, caches. Then it gently trims idle apps’ memory without closing them. When that’s still not enough, Free more closes apps you haven’t touched in 10 minutes. It never closes anything that hosts or serves, tunnels, dashboards, wallpaper engines, terminals, editors, coding assistants or YUKI, and only with a normal close. It tells you what it closed and how much it freed, and offers Reopen. It also shows what’s holding your graphics memory.

Room on disk for downloads. Clear temp files, old runtime backups, download caches and abandoned half-downloads. Your deleted chats and anything still downloading are never touched, and removing a whole model always asks you first.

Host dex inference for chat only. Your PC can now lend chat to dex inference without the 3 GB picture safety check. It’s only needed when you also lend pictures.

Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Your hand only holds cards that can run. If a card can’t start right now (it won’t fit, it isn’t installed, its app is signed out, or its key was turned down), it goes back to your deck with the reason, and Yuki offers the best card that does fit: “Use … instead”. A sleeping brain that fits stays in your hand, and a card that’s only briefly unavailable (a rate limit, a check still running, the pool reconnecting) stays put.

The free board. A new deck tab lists every free model side by side: which providers offer it free, and what’s left today. A “Put in” button is next to each one, plus the providers you could connect for more.

Free models, sorted. The big ones with vision come first; the rest are behind “show more”.

Yuki’s studio. Picture, video and music models are wide studio cards in their own row. Yuki uses them when you ask, so they stay in the deck.

The radar banner clears itself once its source is connected.

Also in this beta

• The groundwork for licensed updates of the paid app (not switched on; this beta updates as before).

The first step toward 0.3. Betas only reach PCs with Settings > Updates > Beta turned on.

What’s new

Every card’s settings, on its back. Flip a card to set what it supports:

• Signed-in coding apps: model, thinking effort, permission mode and sandbox, as each app offers them.
• Local brains: thinking, temperature, context size and the finer sampling settings.
• Cloud models: whatever the provider offers, such as reasoning effort, thinking budget, temperature and longest answer.

Thinking effort now reaches the model for real. Only you can change permission modes and sandboxes; Yuki can’t, not even by undoing a change.

The right context number from the start. A local card no longer shows “1M” after you switch to it. It shows its own window straight away.

“Download me this model.” Paste a Hugging Face link, or ask Yuki. YUKI reads the model’s files without downloading them first: what it is, how big, dense or mixture-of-experts, and how much memory it was trained with. It then picks the best file for this PC and the engine that runs it best, and tells you why in one line. A card lands in your deck when it’s ready. Picture, video and music models arrive as studio cards. A model that can’t run here says why and suggests the nearest one that can.

Hosting dex inference: when the pool turns it on, an answer your PC was writing can carry on on another host if your PC sleeps or closes mid-reply.

Known gaps in this beta

• The rest of the deck refresh (only runnable cards in the hand, the free-model board, studio cards in their own row) comes in the next beta.
• Auto-tuning a newly downloaded model for your PC comes later.

What YUKI is

YUKI is an AI agent that lives on your own Windows PC. You talk to Yuki and she does the work on the machine: files, the shell, the browser, your desktop apps, pictures, video and music, and other AI models when a job needs them. Her brain can run entirely on your own graphics card, and your chats stay on your PC unless you pick an outside model.

0.1 only ever ran on the developer’s own PC. 0.2 is the first release anyone can download.

What’s new in 0.2

The hand and the deck. Models are cards now. The cards in your hand are the ones Yuki works with; everything else she can reach (local models, outside providers, signed-in apps) waits in the deck. Drag a card into your hand and it wakes up: a local model loads, a key or a sign-in gets checked, and a card that can’t start comes back out with the reason. Alt+1 to Alt+9 switch the chat to that slot, and each card keeps its settings on its back. Yuki can manage the hand with her own tools, but unloading a model to make room always asks you first, and so does a paid provider that isn’t in your hand.

The free radar and new providers. YUKI keeps a list of free model sources: what each one gives, its limits, and its catch (like “logs your chats”). That list and its limits come with app updates; OpenRouter’s public model list is the one part that refreshes on its own, once a day. When a new free model turns up there, a banner tells you, at most once a day. Connecting a provider opens a card with the exact steps to get a key and a box to paste it in; a good key unlocks that provider’s models into your deck, free ones first. New in this release: OpenRouter, opencode Zen, NVIDIA NIM, Cloudflare Workers AI, Z.ai, TokenRouter and Token-Router, your own keys for Anthropic, OpenAI, Google, Groq, xAI, Perplexity, Fireworks, Cerebras, Hugging Face and Together, and a local model server on your own PC. Paste a key and YUKI works out which provider it belongs to.

A local brain that fits most PCs. First run reads your hardware, picks the best brain in the catalogue that your PC can actually run, and tells you why in numbers when none fits. A 24 GB card gets a 27B model (or a bigger one with 64 GB of RAM). Smaller cards get a compact mixture-of-experts model that keeps what doesn’t fit on the card in RAM, so it runs on a 6 GB card with 16 GB of RAM. On a 6 GB card it needs about 9 GB of free RAM beside the card, and it works best with heavy programs closed. Measured inside YUKI on the 6 GB / 16 GB test laptop with ordinary apps open, it writes about 17 tokens a second on short replies and about 13 on long ones (about reading speed), and reads prompts at about 700 tokens a second. With more free RAM, the first reply after a wake and the re-read of her instructions get quicker. First run sizes it against what’s free at that moment, and when your PC is busy the brain wakes a little smaller and tells you so instead of refusing. The first reply after it wakes takes about half a minute, because it reads its weights from disk as it goes; after that, new chats start in about a second. On a busy PC there may be no room to set YUKI’s instructions aside while Yuki names a new chat, so the message after she names one reads them again first: about 8 seconds on the test laptop. A PC that can’t run a local brain can start on a free cloud brain instead.

Compact mode. YUKI can fold down into a small square with Yuki in it that stays on your desktop, right where the window’s corner was. Hover it to type. The latest reply shows up with its pictures and files as tiles you can drag straight into other apps, and everything that’s running floats below her as cards you can open or stop. A small pill under her switches chats.

Voice. Turn on the speaker and Yuki talks. Speaking is a tool she uses: she writes a short spoken line for each reply (she won’t read code or long text out loud), and push-to-talk lets you answer by voice. Her voice runs on the processor. Her ears are a listening model, with a second one for Japanese, Chinese and Korean.

Signed-in apps. A coding app you already use can run inside a YUKI chat on your own sign-in and plan. YUKI keeps the transcript, picks the app’s own session back up next time, and lends the app her tools (pictures, video, the local model, the browser) through a locked server on your PC. She never sees your password or session token. Other coding apps have routes too, but haven’t been tried on a real install yet.

The dex account. Optional; YUKI works the same without it. Sign in with Discord, GitHub or Google (no passwords) for end-to-end encrypted sync of your chats, prompts and settings. It’s also how dex inference, the shared model pool, will work. The pool opens to friends first, so it isn’t open to everyone yet.

The web app’s Your PC mode. The web app has a Your PC mode: a remote for the YUKI on your own PC, paired like the phone, with your hand, your local models and every tool still running at home. It’s live at https://dex.place/yuki/app/.

The update card. From 0.2 on, YUKI updates itself. It checks at launch and every few hours, downloads only what changed, and puts a card at the top of the chat when the update is ready. Restart waits for a reply that’s still coming; Later lasts until the next launch. Settings > App > Updates has Check now and the stable or beta channel.

Mascots and design. A new look from top to bottom: pixel-art faces for Yuki and every helper type, with expressions and idle blinks, a pixel icon set, springy motion, and tiny sounds made in code (with a volume slider that goes to mute). The orb is redrawn so it can grow and shrink with compact mode.

Known issues

• Films can crawl and time out on a 24 GB card when other apps hold graphics memory. The video model needs about 21.5 GB, so anything else on the card pushes it into system memory, where it slows to a crawl. The fix is planned for 0.2.1. Until then, close other programs that use the graphics card before a film. Setting NVIDIA Control Panel > Manage 3D settings > Program Settings > CUDA - Sysmem Fallback Policy to Prefer No Sysmem Fallback for the python.exe that runs ComfyUI (in YUKI’s runtimes folder) also helps.
• The phone app hasn’t been tested on a real phone yet.
• The small brain can miscount exact facts. The 3-bit model for small cards can get numbers, counts and dates wrong. Check anything exact it tells you, or give that job to a bigger or outside model.
• Yuki’s bigger graphics-card voice can’t be installed from the app yet. Voice runs on the processor.