Yumina

Updates

Regex: replace or remove what the AI keeps saying

Settings > AI Config has a new Regex section. Fill in Find and Replace with, or leave it empty to remove; tick Regular expression to write a real regex. Each rule can apply to the AI's replies or your own, on screen or in what is sent to the AI. With Sent to the AI on, the AI stops seeing…

Light mode for play

Choose Dark or Light in Settings → Display / Language → Play appearance. Dark remains the default.

Faster, cheaper default models

The quick-select row adds DeepSeek V4.1 Flash 1017 and MiMo V2.6 Pro: the first answers almost instantly and costs the least, the second writes steadily and rarely needs a reroll. Hy3 and MiniMax M3 leave the row and stay in the full model list. Mix mode's default blend is now MiMo V2.6 Pro,…

Turn one picture into a 3D model and place it in your 3D scene

The Atelier has a new studio, AI 3D models : pick a picture and get a textured 3D model, saved to your asset library as a .glb. Turn it around on the page, or download it. Choose the generator (Hunyuan 3D 3.1 Pro by default, 630 Mushies) and the detail: "High" for things seen up close,…

New model: Gemini 4 Hydrogen

Gemini 4 Hydrogen: for Gold members, in chat and Studio. Clean prose that keeps to length, and it reads images. Its rate is lower than Gemini 4 Argon, but it always thinks, so it costs about the same as Argon on Low and takes a little longer to start.

A card's interface can now draw its own pictures

Custom interfaces get api.images.generate : a player clicks a button and the card draws a picture, such as a newly met NPC for a gallery or the current scene for a panel. It is drawn and billed like a reply's picture. Before a card's first picture, Yumina asks the player, and draws only…

Mechanic packs moved into the Resource market, and its front page is cleaner

The four mechanic packs, Affection, Survival meters, Random events and Status panel , are now official packs in the Resource market, each with a cover that shows the numbers and rules it adds to your card. Mechanics is the first tile on the market's front page. Every place on the canvas that…

"Modules & Prompts" is now the Resource market, and prompts have a new look

On Discover, Modules & Prompts is now the Resource market . It opens on bundles, with prompts beside them. Prompts have a front page: the most installed one on top, with the opening of the prompt itself right there; then the newest, the most recently verified, and tiles for Style, No speaking…

Kimi K3.1 Pro in chat, Kimi K2.6 on free accounts

Kimi K3.1 Pro is available in chat and Studio for Platinum members. Kimi K2.6 no longer needs a membership; free accounts can pick it. Kimi K2.5 shuts down on November 18. From then on, chats that used it switch to Kimi K2.6 on their own. Hermes 3 405B and Euryale L3.3 70B are gone. Chats that used…

Run open models from Hugging Face on your own computer

Run it on your own computer in Settings has a new Hugging Face models section. Paste a model link to see the versions you can download, their size and the VRAM each needs, then hit Copy command and run it in a terminal. Original weights from a lab can't run as they are, so pasting one lists…

Look, code and templates are easier to find

The Player interface block has Edit (each part's colours, fonts and corners) and Code (this card's interface code). Type "theme", "code" or "template" in the canvas search and it takes you straight there. Front End Code is now Interface code ; Export Bundle is…

A clearer canvas

New cards open straight on the canvas. Each block reads from the opening down: lore, then variables and behaviors, and the AI that answers at the bottom. The Player interface block has Switch interface : pick a look and it is on, and Undo takes it back. Scroll down and interfaces from the Resource…

Design a character's own voice from a description

The voice picker has Design a voice with AI under Custom voices . Describe the voice in a sentence, such as "a woman in her twenties, low and a little husky, speaks slowly", then click Generate for 3 candidates. Play them and click Use on the one you want. It becomes this character's…

Spend tip earnings on mushies

Under Mushie Packs , switch the payment to Tip earnings to buy packs with the tips you received, with 5% more mushies that never expire. It works wherever you live, with no payout account. Convert all turns the rest into mushies. Setting up payouts now asks where you live first, so creators outside…

Mark library games Playing, Finished or Dropped

Open a game in your library and set its Status to Playing , Finished or Dropped . Anything you've played counts as Playing until you mark it. The library has Playing , Finished and Dropped filters, and each card shows its status.

Prompt switches are remembered per world

Prompts you switch on or off in chat under + › Prompts now apply to that world only, and stay that way when you come back to it. Prompts that only suit some worlds, like a writing style, no longer need switching by hand each time. Prompts you never changed in a world still follow the switches in…

Choose how much the recap folds, delete chapters, and chapters numbered right

The Memory panel has a new Auto-fold line: how many tokens of word-for-word history there are now, and the point where the recap starts folding on its own. Only when the history is about to fill story memory does the older half become chapters, so with a large story memory a long stretch goes by…

Use someone else's interface and have the assistant fit it to your card

The market has a new Interface category: complete interfaces other creators made, such as status bars, inventories, relationships and maps. Pick Interface in the market's type filter to see only those. The canvas's + menu has a new Use an interface template , which opens the resource market…

See the Studio assistant's plan and add to it while it works

When the assistant works through a multi-step task, a Plan 2/5 line appears above the input box with the step it's on. Click it to see every step. The line goes away when the task ends. You can type and send while the assistant is working, without stopping it. It reads your message before its…

Look around a bundle's canvas before you add it

Click a bundle in the canvas's Bundles panel and the preview now has Screen and Canvas buttons at the top. Canvas shows what is actually in the bundle: each entry, variable and behavior, the way it will sit in your card. Open large opens a big window with the bundle's canvas on the left and…

Long stories remember more: the recap, pinned notes and the Memory panel

Older scenes that no longer fit in story memory are now folded into a recap, chapter by chapter, and sent to the AI with every reply. We tested it on real long stories: answering questions about the early story from the recap alone, the new recap got about half right, the old summary under one in…

Writing in a canvas row no longer gets interrupted

While a row was open on the canvas, a click on the bare part of a nearby block (like the Context block right under the editor) closed the row you were writing in and opened that block instead. Now a click on empty space leaves your row open; click a row or a block's title to switch. Clicking…

Drag bundles onto the canvas, then set them up and update them there

The canvas toolbar has a Bundles button right of + Add . It opens the bundles panel on the left of the canvas. Click one and a preview opens beside it, with its screen running and a list of the entries, variables and behaviors it would add. Drop a bundle on empty canvas to add it to the card. Drop…

Cards with their own screen open faster

Cards that bring their own screen, like simulators, maps and mini-games, used to rebuild that screen on your phone or computer every time they opened, which cost up to two or three extra seconds on a phone for the big ones. Each card now does that once, the first time anyone opens it, and opens…

Play a card before rating it, and say why for 1 or 2 stars

Stars on a card's Reviews tab now open once you've played that card. Comments work as before, played or not. Pick 1 or 2 stars and a list of common reasons appears under them, such as Not as described , Too little content or Characters feel off / OOC . Pick as many as apply; most save with…

Connect DeepSeek to your cards

Connect your AI in the Studio top bar has a new DeepSeek tab, so DeepSeek Harness (dsh), DeepSeek's own agent, can read and edit your card. Click Make token and the dialog hands you a filled-in yumina.yml plus the one command to run. You'll need Node.js installed. The token works on the…

Doubao Seed 2.1 Turbo joins chat

ByteDance's Doubao Seed 2.1 Turbo is now in the model picker, free accounts included. It answers without a thinking pass, so replies usually land in two or three seconds.

Video uploads for animated covers

Upload an MP4 or WebM, or choose a video from your assets as a cover. Animated cover sources can be up to 64 MB and 60 seconds. They are optimized after upload, with playback controls available while cropping.

Comment from anywhere in a thread

A persistent comment bar opens the reply editor. Minimizing keeps your draft; posting keeps your current page and reading position.

Cancel chat image generation

Use “Cancel generation” while drawing or redrawing. Cancelling refunds any charge for that picture and keeps the original image when redrawing.

Opus 5.5 and other new models in Studio, GLM 5.3 in chat

The Studio assistant can now run on Claude Opus 5.5, Claude Sonnet 5.5, Gemini 4 Argon, GPT-6.1 Sol, Grok 4.7, Kimi K3, Qwen3.8 Max, GLM 5.3, MiMo V2.6 Flash, DeepSeek V4.1 Flash 1017 and GPT-6 Luna. Studio's model list now groups versions of one series at the same price into a single row, like…

New models, and one row per series at the same price

Added Gemini 4 Argon, Claude Opus 5.5, Claude Haiku 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, GPT-6 Luna and Grok 4.7. Haiku 5.5 and GPT-6 Luna work on free accounts. Versions of one series at the same price now share a row; tap a version number in the row to switch. The Claude Opus row, for example,…

Pictures in card descriptions

When writing a card's description, click the image icon above the box to Upload from computer or Choose from assets , or paste or drag a picture straight in. The picture goes where your cursor is and shows full size on the card's page, so each character can have their own. In the box it…

Story illustrations now use Anima

Story illustrations are now drawn with Anima : pictures arrive faster, and group scenes follow the story more closely. The style looks a little different. If a picture misses, use Redraw or Redraw with a note . For creators: Anime and Limitless Anime in Custom generation have an Anima version.

Update log

New features and fixes are listed here. Open Notifications and tap the button left of the gear. Tap a date to copy that update's link.

Improvements and fixes (Sept 29–Oct 10)

Describe & create and video have more models to choose from. The save list has Select and Select all to delete several saves at once. You can choose the order of public works on your profile. Library has Save images for pictures saved during play. For creators: the Creator Guide is rewritten…

Player interface templates

In Player interface , open Template to use one of 12 full-screen layouts, with variables and buttons already wired. Change text, colors and positions right on the preview.

The canvas editor

Turn on Canvas at the top left of the editor to see the opening, lore, variables and player interface on one board, and click anything to edit it. Turn it off to go back to the usual editor; nothing is lost.

A new Discover

Discover is sorted into Anime, Games, Roleplays and Novels & Literature, with tabs for Following, Bundles and Prompts. The feed doesn't repeat cards as you scroll, and tells you when you've seen everything.

Story illustrations

Tap Draw this scene under a reply to turn that moment into a picture. Not right? Use Redraw , or Redraw with a note to say what to change. For a picture every turn, turn on Illustrate every reply under Settings › Display / Language › Story illustrations.

AI video generation

AI Generate has a Video tab: a sentence or a picture is enough for a short clip. Creators can put clips into a card, for example as a character's moving portrait.

Art styles and 4K in custom generation

Custom generation offers styles such as Anime, Realistic, 2.5D, Ink Wash and Western. Quality can be Standard, HD or 4K, and a finished picture has a 4K version button.

Voice readout and voice input

Turn on Voice readout in Settings › Display / Language, then tap the speaker next to a message; the narrator and each character have their own voice. Voice input : hold the mic and talk instead of typing.

Improvements and fixes (Sept 15–28)

You can upload or paste images in chat. The model picker has search and favorites. Discover has Prompts to install prompts other players share, and Settings › Prompts lets you bind one to each model. Daily check-in is now daily and weekly quests, collected automatically. For creators: AI Generate…

Play with models on your own computer

The model picker has This computer : with Ollama or a similar app installed, click Check this computer to play on your own models without using Mushies. Desktop only for now; a setup guide sits beside it.

AI image generation

Open AI generation on the Create page, go to Image , and write a sentence to get a picture. Attach a reference picture to have it changed the way you describe. Choose the aspect ratio and size; pictures are saved to your Assets.

Improvements and fixes (Sept 1–14)

On phones, the top bar tucks away as you scroll a story. Switch to another app and back, and the chat stays where you were reading. Changing your persona in Settings applies to all your chats. Tag filters now include your own works and ones you've favorited.

Friends and a floating message window

Follow each other and you're friends; in multiplayer games you can invite a friend in one tap. Messages can open as a floating window, so you can keep browsing while you chat. Similar notifications are grouped into one row, with filters for follows, replies and likes.