Release Notes

2.0 started the engine. 2.1 hits the road.

While others are still figuring out how to start the engine, we've already floored it.

New features

• Sumi Academy: Don't just watch the film, see how it was made. Curated official works in one place, with the final cut playing on hover. Tap “See how it was made” and everything is laid bare, from the first sentence to the final cut. Like it? Copy it in one click and take it with you.
• Process replay: Rewatch every step, while others hide even the process. Community works can make their creation process public, with a segmented progress bar you can jump anywhere on. How others thought, revised, and rendered, all laid out in front of you.
• Proactive Sumi alerts: You go do your thing, it calls you back. When you need to make a call or a turn finishes, system notifications, the tab title, and alert sounds all kick in. Two sounds for done and awaiting confirmation, so you can tell at a glance, by ear. Others make you watch the progress bar; Sumi has learned to nudge you.
• Interrupt anytime: Say it when you think of it, no waiting. While Sumi works, you can add new requests, and a message that hasn't been delivered can still be withdrawn. Others make you wait in line; Sumi lets you cut in anytime.
• Shot plan check: Check first, then spend. Before generating video, Sumi checks segment by segment against the shot plan, and the payment confirmation card flags anything that doesn't match; shot plans in the workspace display as tables. Trying to waste your money? Not a chance.
• New Kling capabilities: Four of them, all at once. Kling text-to-speech makes digital humans talk; advanced lip sync supports multi-person videos, detecting faces first and then letting you pick one; single- and dual-subject video effects work out of the box; custom voices can go into video generation; and elements from your element library can serve as Kling subjects. Others squeeze it out drop by drop; we serve the whole pot.
• More models: Six more, the arsenal expands again. Claude Opus 5.5, Claude Sonnet 5.5, GPT 6 Luna, GPT 6 Sol, GPT 6.1 Sol, and Nano Banana 2.1 arrive. Others are padding the numbers; we're shipping new ones.
• More official Skills: Skills aplenty. Nine new social-media planning and operations Skills, 16 English Skills and the “English Creator Toolkit” collection, plus three advanced Skills for creative roadmaps, promo development, and model workarounds. Others hand you tutorials; we hand you the whole toolkit.

Improvements

• Connect an Agent, in two steps: Copy a passage to your favorite Agent, connect your account, and start creating. Desktop clients, command lines, and the web, mainstream apps all covered. Others are still teaching you how to configure; we've made it copy-and-paste.
• Official Skills, fully revised: Twenty-plus official Skills rewritten along the complete creative pipeline, from script and shot plan to voice-over and editing, for smoother handoffs and steadier output. Others patch things up; we rewrite.
• Model options, all in: The GPT Image series supports custom and automatic sizes; Claude's thinking can be switched on or off; Gemini image generation accepts PDF input; GLM supports web search. Others give you what they give you; we let you choose.
• Sumi guidance, just enough: The first time you reach an easy-to-miss entry, Sumi drops a word; once said, it won't bother you again, and you can turn it off or replay it in settings. Others can't teach; we say just enough.

2.0 started the engine, 2.1 pulls out.

See the process, keep the pace.

Sumi 2.1, a new era “Sets Off”.

Others are still catching up; we're already on the road.

This is not an ordinary update. This is a generational shift in how films get made.

The all-new Sumi 2.0, a new era, takes AI from “helping you write a little” straight to “assembling an AI film crew for you.”

New features

• Introducing Sumi, a professional film Agent: One sentence, and the cameras roll. Share your idea and Sumi sets up the project, writes the script, lays out the canvas, produces the assets, and cuts the film; it asks questions, proposes a plan first, and carries the work all the way through. You bring the inspiration; it delivers the final cut.
• The all-new Harness architecture: An Agent engine evolved to legendary status. The new engine behind Sumi keeps long tasks connected, keeps big projects on track, and runs multi-step creation in one smooth flow. Even complex film workflows move as smoothly as a director calling “Action.”
• Full-featured MCP is live: Platform capabilities, fully remote. Connect the platform to the AI apps you already use — generate, edit, manage assets, and run canvas nodes; everything the platform can do can be called remotely. It is not just an interface; it is an entire creative factory in your pocket.
• Full support for open-source Skills: A skills ecosystem, power-ups in one click. Compatible with the open-source Skill format — import, upload, and install in one click; official skills and Apple Box skill collections work out of the box, and your favorite workflows can be saved as Skills too. The more you use it, the stronger it gets; the stronger it gets, the faster you go.
• All-in-one resource management: Every project asset, under control. Scripts, canvases, edits, assets, and conversations live together by project, files can be browsed level by level, @ references any element in one tap, and storage and usage are clear at a glance. Clutter? Not here.
• Team features are live: One team, one studio. Team spaces share projects, canvases, and edits, character cards are shared within a project, and teams manage their own seats and subscriptions. No more passing files around — shoot together on the same set.
• A more open community: Inspiration flows, copied in one tap. Projects, scripts, canvases, edits, and characters can all be shared to the community, and others can copy them into their own projects in one tap; finished work can get a share link. Great ideas are no longer islands — the whole community is your asset library.
• Smart editing: Professional post-production, right in your browser. Editing projects can be created on their own; keyframes, masks, keying, curve grading and curve speed ramps, dozens of transitions, styled text and text animations, beat sync, and multiple timelines side by side are all here. What used to take several apps now takes just Sumi.
• More models: 32 top-tier models — the arsenal arrives all at once. Claude Fable 5.1, GPT 6 Astra, Gemini 3.8 Flash, Grok 4.7, Veo 3.1, Seedream 5.0 Flash, Suno v6, and more — 32 models in all. Inspiration never lacks weapons; creation comes fully loaded.

Improvements

• Privacy protection, the safest platform: Your work is private by default. Your work is yours alone to access. Peace of mind, maxed out.
• Brand upgrade: A fresh look across the whole site. Supports light mode and can follow your system setting. Better looking, easier to use, more like the future.
• Performance improvements: Fast, faster, and a little faster still. The homepage opens faster, and the canvas and editor feel more responsive. Your creative rhythm should never be interrupted by a loading bar.
• Pay for actual usage: Transparent down to every cent. Sumi conversations are settled by real usage, unused prepaid credits are returned on the spot, and the bill shows exactly what was charged and what was returned. No overcharging, no confusion, no paying for thin air.

Sumi 2.0 is more than a version number — it is a new era for AI filmmaking.

If 1.0 was a tool, 2.0 is an AI film crew that is on call, always online, and takes you from script all the way to final cut.

One sentence, from inspiration to final cut. The all-new Sumi makes its grand debut.

New features

• Canvas AI assistant (preview): describe what you want in one sentence and the assistant lays out nodes, picks models, and runs the whole chain on the canvas — every step visible, stoppable at any time, with credits spent shown at the end of each reply. Now in beta preview, with some abilities still being polished
• Creative skills, open to all: use official skills out of the box, or package your own workflow as a skill and share it, with favorites, popularity, and showcase pieces at a glance
• Engine in three tiers: the assistant's engine comes in Nimble, Versatile, and Flagship tiers with free model choice within each, synced across the chat box, the /model command, and the engine picker; if a run fails, retry with a different engine
• Attachments in chat: drag images and files straight into the conversation — they adapt to what your chosen engine can read
• Creation preferences: pick preferred models for generation nodes, save as many as you like, and the assistant honors them when choosing models
• A refreshed homepage: the homepage opens straight into a full creation box — one sentence takes you to the canvas; pick a skill and bring your idea along
• Seven new Suno models: Suno v3, v3.5, v4, v4.5, v4.5+, v5, and v5.5 all arrive, with a dedicated lyrics field in custom mode
• Three new Seed Speech models: Seed Audio 1.0, Seed TTS 2.0, and Seed ASR 2.0 Flash cover speech synthesis and speech recognition in one go
• Thai interface: the whole site now speaks Thai — one more language in the family

Improvements

• Results right in the chat: generated images, videos, and audio preview as cards inside the conversation — no need to go back to the canvas to find them
• Context caching, your call: multi-turn context caching is now a user-selectable option, off by default
• A refreshed music player: lyrics scroll with playback and the current line lights up, tracks from the same batch switch freely, and starting one track on the canvas pauses the rest
• Better on mobile: app-wide motion and narrow-screen layouts tuned, with the canvas opening as a read-only preview on phones
• Clearer voice management: cloning targets grouped into voice and music, and MiniMax voices can be auditioned before you pick

New features

• Seven new Claude models: Fable 5, Opus 5 / 4.8 / 4.7 / 4.6, Sonnet 5 / 4.6 all at once, and every one of them reads PDFs directly
• Three new DeepSeek V4 models: V4 Pro, V4 Flash, and V4 Flash Vision with vision input
• Three new Grok models: Grok 4.6, 4.5, and 4.3
• New model Grok Imagine Image 2.0: one more brush for image generation
• New model Wan 3.0 Video: standard and turbo tiers go live together, for both text-to-video and image-to-video

Improvements

• Smoother asset flow: drag assets into the generation bar as references, move assets into any project, and cut a timeline clip straight back to the canvas as an asset; the canvas left rail is now a template library plus an asset drawer

New features

• Independent AI Audio mode: voice synthesis and text-to-music become their own creative mode with a more focused interface; the input bar gets translate/optimize/restore buttons
• New batch creation modes: Stack (album) and Angles ("best angle set") turn one generation into a full set
• Community sharing and one-tap reuse: share your quick-generate results to the community in one tap, view generation parameters on any post's detail page and reuse them instantly; the homepage community section now splits into "Editor's Picks" and "Community Picks"
• Assets and project management overhaul: organize by project with nested folders, a docked generation bar on the assets page, and multi-select send-to-canvas from the my-assets panel
• Unified reference picker: adding a reference in the generator or picking assets now goes through one asset-library dialog with inline upload; the playground also gains reference-driven video generation
• New credit calculator on the pricing page: pick a model and see how much you can generate per month, with the top banner now showing real per-second pricing
• Unified account navigation: settings, subscription, redemption codes, usage, queue, team, tickets, and notifications now live in one place, with usage trends shown as real charts
• Expanded model lineup: a digital-human model, the MiniMax H3 voice-and-video family, and Seedance 2.5's 1080p tier are all live, plus transparent background and layer decomposition for image generation
• Smoother canvas navigation: task history can jump back to the canvas and land on the exact node, node outputs page through past generations with a total count, and ungrouped node chains can run in dependency order with one click

Improvements

• Unified generation bar across modes: image, video, and audio generation bars now share the same look and interaction, with clear pricing and playthrough guidance
• No more guessing on failures: generation history supports one-tap retry, shows the failure reason clearly, and refunds credits when due
• Cleaner plan cards: removed the "max generations per month" overview in favor of an in-card limited-time promo section
• Clearer renewals and upgrades: renewal dates are now labeled distinctly, Alipay plan upgrades get a confirmation dialog, and redemption-code upgrades warn about credit reset with an itemized confirmation
• Refreshed footer and help links: dead links cleaned up, now pointing to real support channels
• Snappier overall: canvas, batch submission, and queue concurrency and responsiveness improved across the board, with app-wide motion, frosted glass, and press feedback polished to a consistent standard

New features

• Node annotations: leave comments directly on canvas nodes, with unread markers and titles that follow your language automatically
• Inline canvas clipping: video clipping now lives right in the canvas toolbar, paired with a brand-new player, so you never leave the page

Improvements

• Smoother connections and alignment: dropping a connection on empty space creates a node right where you dropped it, and selected edges now use a continuous gradient outline
• More accurate pricing: portrait verification and image-generation holds are calculated against the actual tier, so balances and remaining quota no longer read higher than they are
• Uploads become asset nodes: dragged uploads now create their own asset node, making the source clearer

Fixes

• Sturdier video result cards: the player fills its full height after a refresh, and cards are no longer capped by a fixed height
• More reliable task status: tasks with missing output no longer get marked done early, so results don't get stuck unreachable
• Small polish: the side drawer no longer flashes blank on slow networks, switching to a model without storyboard support now confirms before clearing it, and new video nodes created from a video node inherit its reference settings

New features

• Canvas recycle bin: delete without fear — restore or clear anytime
• Annotation trio: sticky notes, brush, and five-color markers keep your ideas clear
• Smoother connections: a connection menu, alignment guides, and batch connect — drag assets straight onto the canvas as nodes
• Ukrainian joins the list: the site now supports 14 languages
• Alipay buyout for teams: pay once, no subscription needed for new team purchases

Improvements

• Multi-shot editing moves into the generation bar: first and last frames are set, creation flows better
• Prompt optimization gets smarter: matches your target model and language automatically
• A calmer pricing page: plans scroll horizontally instead of crowding together
• Version number lives in the footer: one tap to the release history

Fixes

• Multi-shot false warnings gone: every filled shot is recognized, and Kling balances durations correctly
• Long videos stop stalling: the timeline holds steady past 30 seconds
• Canvas interactions feel sharper: sticky notes, multi-select, dragging, and placement all smoothed out

New features

• Team collaboration upgrade: invite links and join approval make collaboration clear at a glance; copy any personal canvas to a team in one click
• Self-serve plan management: upgrade tiers or add seats right from the team page
• Team asset library: assets and storage shared across the team
• Buy memberships outright with Alipay: pay once for monthly / quarterly / yearly plans, first-purchase discounts still apply, and subscribers can buy up to a higher tier too
• Smart References: a new reference system that keeps results closer to what you have in mind

Improvements

• Redemption codes copy in one click, and bulk downloads show packaging progress
• Team pricing details are clearer, and review results now arrive as soon as they're ready

Fixes

• No more horizontal overflow on mobile, and collaborative edits are protected more reliably
• Team viewers now open canvases in proper read-only mode

Improvements

• Upgraded release pipeline: new features and fixes reach you faster and more reliably
• Complete release history: starting with this version, every update is fully documented here

New features

• New payment result pages: dedicated pages for success and pending payments
• In-app checkout: buy credits without leaving the site
• Frame capture: grab the current, first, or last frame as an image node in one click
• Generation settings on public canvases: select a node to see its prompt, model, and parameters

Improvements

• Upscale FPS can be left empty to follow the source, with pricing tiers clearly labeled
• Invalid emails get an instant inline hint, and AI MediaKit models show their brand icons

Fixes

• Fixed the parameter popover blocking canvas panning and the checkout dialog being covered
• Qwen image-to-image keeps your aspect ratio, and team plan prices are no longer clipped

New features

• New models: Qwen Image 3.0 and Wanxiang 3.0 arrive, with a full parameter refresh for the Seedance family
• Signup gift: new users get bonus credits and their real best first-purchase discount
• Stackable limited-time discounts: model discounts plus sitewide offers, with real pricing shown up front
• Zoom into results: browse and zoom through result cards
• Singapore Dollar added as a display currency

Improvements

• Video duration is now a slider with available seconds at a glance
• Models with a recommended effect carry a glowing badge, and the workspace was redesigned for bigger previews
• Community posts can be edited after publishing, and login now leads with password sign-in

Fixes

• Model filters are remembered when you return from a detail page
• Video exports no longer get occasionally stuck, and discount labels are more accurate

New features

• A brand-new video player: faster loading with quality switching, across community and detail pages
• Storyboard editing upgrade: edit text in place, drag across tracks without losing position, plus audio tracks and an end-of-film marker
• Web search: visual understanding tasks can now search the web, including text-to-image and image-to-image search
• Ambassador program redesign: an apply flow plus a promotion dashboard, with discounted prices shown up front

Improvements

• Canvas nodes get a unified control bar, and text nodes are editable as soon as you select them
• Canvas performance: off-screen animations pause automatically for smoother scrolling

Fixes

• Fixed deleted storyboard clips reappearing and end-of-film flicker
• Audio nodes that fail to load can be retried, and failed post covers show a clear message

New features

• A new asset workbench: name your assets, search by name, and download a ZIP — all in a full-screen page
• Invite contest: multi-wave seasons, a $3,000 prize pool, and a Top-10 leaderboard
• Redeem codes in place: right inside the credits panel
• Copy docs as Markdown: handy for handing off to AI assistants

Improvements

• Generating cards show a flowing water-light effect, and the task queue is now paginated
• Social share images use a canvas collage layout that shows off your work

Fixes

• Fixed nodes stuck in queued state, and long-unresponsive tasks now auto-cancel and refund
• Downloaded file names match the titles shown on cards

New features

• Mobile canvas: touch panning, a slimmer generation bar, and a collapsible top bar
• Multi-track audio: multiple tracks, crossfades, and transitions
• AI watermark: on by default, with your own image and text as an option
• Email notifications can now be chosen by type

Improvements

• The pricing page was redesigned around monthly plans, and inpainting brushes are more precise
• The model dropdown shows brand icons, sitewide motion got finer, and share previews were polished

Fixes

• Fixed timeline audio tracks staying silent and trim handles drifting while dragging
• Images are re-signed before copying to the clipboard, so copies no longer fail on expired links

New features

• Portrait library: register a portrait and use it in generation right away
• Character library refresh: all 150 preset characters got face avatars and tag cards
• Aerial camera-move template: 3 real landmarks × 3 camera moves, ready out of the box
• AI assistant upgrade: paid analysis subtasks always ask for approval first, with transparent pricing

Improvements

• Camera control now scroll-snaps vertically, and finished images show at full resolution
• The whole site adopted a liquid-glass style, with model icons completed and sorted

Fixes

• Fixed occasional failures in photo enhance and broken community video covers
• Billing and permission hardening for more reliable charges and refunds

New features

• Telegram / Discord bots: generate right inside your chat apps, and sign in with either account
• Instant subscription upgrades: pay only the difference, effective immediately
• Login redesign: a rebuilt sign-in dialog with email verification for password sign-up
• Photo enhance: image upscaling relaunched under a new name

Improvements

• The compose timeline supports reordering, cross-track inserts, and trim handles
• Dialogs use our own design sitewide, credits get a consistent lightning icon, and history items can be favorited in one click

Fixes

• Fixed nodes from the “+” menu not auto-connecting and subscription credits occasionally not granted
• Text-generation results can be edited directly, and the AI-write bar can be collapsed

New features

• Rebuilt membership plans: unified USD pricing with clearer benefits
• Universal interpreter: read images, video, audio, and text in one click
• Assistant confirm / auto modes: choose how hands-on you want to be
• Daily login credits now refresh every day

Improvements

• The blog moved under Community with an autoplaying featured carousel
• Blocked images get a clear message, and public pages load noticeably faster

Fixes

• Fixed results vanishing from the canvas on refresh
• Cancellation refunds are more reliable, and duplicate or missing credit grants are fixed

New features

• A big model expansion: more image and video models to choose from
• Fairer batch generation: settle by actual output, with automatic refunds for the rest
• Better video understanding: reads the original footage directly for more accurate results
• Team-exclusive model quotas

Improvements

• Rename canvas node titles with a double-click, and selected edges show a meteor-trail effect
• Storyboard generation uses a cleaner one-row-per-shot layout, and timeline preview eliminates stutter

Fixes

• Fixed occasional failures chaining reference images into image-to-image
• Text nodes no longer silently charge for media inputs

New features

• Canvas interaction upgrade: marquee select, wheel panning, ⌘ zoom, and smoother edges with @-mention assets
• Director stage refresh: real skeletons with 43 animation presets
• Template library: a home module and a dedicated page to start creating in one click
• Team top-up pool with full transaction history, plus profile avatars

Improvements

• The home carousel went all-video, and dialogs adopted a liquid-glass look
• The text editor was upgraded, and context menus and paste were rounded out

Fixes

• Node titles now display in your interface language
• Marquee multi-select no longer expands every node editor

New features

• Tasks under your control: progress, retry, and cancel at hand, with automatic rollback on cancel
• Inpainting: regenerate just the region you select
• Reference-to-video: all 12 video models support reference mode
• Smart describe: auto-generate a description from a reference image
• Voice library / notification center: an official voice catalog, and paginated notifications by type

Improvements

• Empty nodes show creative examples, and video compose supports per-segment mute
• The timeline supports splitting clips and dropping in local files

Fixes

• Fixed Chinese IME breaking words in text boxes
• Multi-image generation no longer overcharges, with clearer retry guidance

New features

• AI canvas: connect nodes to organize your work, with local files becoming asset nodes automatically
• Director stage: a searchable 3D scene tree with multi-camera management
• Storyboards: editable shot durations, stitched into a film in one click
• Video compose: adjustable output with vocal / background separation
• Image upscaling / membership: boost resolution in one click, with your history always in view

Fixes

• Fixed sign-in occasionally not responding
• Notification dates now display correctly in your language