Changelog

Follow each release: new capabilities, fixes and everything you need to know before upgrading.

Published releases, newest first, including stable and pre-release versions. Synced hourly; showing the latest 30 releases.

October 21, 2026
StableLatest release
v2.0.0
- 桌面端接上 MosaelCloud,官网长出一个社区。
- 两个身份,两套登录。 设置里新增「MosaelCloud 账号」:它和本地账号互不相干 —— 云端登录态由桌面后端转发票据、存在这台机器上,密码不落本地库;本地没登录时云端那块也不会跟着掉。
- 社区上线。 /community 是社区首页(插件 / 工作流 / 创意画板三段),三个市场页各自能搜、能筛、能排;官方内容由官方账号(中英各一份)发布,任何人都能上传自己的 .json,/community/upload 是唯一入口。
- 分享详情页能看内容。 工作流和画布画成结构图;画布还能进无限画布里平移缩放浏览(只读)。
- 创意画板能导出导入。 「更多」里导出 .mosael-board.json,清单里导入;导出的就是社区收的那一份。
Full notes and downloads
October 20, 2026
Stable
v1.9.11
- Dubbing has a default now. It was the one capability missing it — transcription, denoise and speech synthesis all had a "use this one by default" row, while dubbing made every caller name its own engine and voice. The default is not a single dropdown but a pai…
- Capability providers are scannable now. One card per capability, with the title row stating the current choice ("Denoise · Built-in", "Dubbing · Not set") so you no longer hunt for the dropdown; "used in" collapses to a single summary line that expands on dema…
- No more button that only explains itself after you press it. The app now knows on first paint whether this deployment is connected to MosaelCloud; when it is not, the button is disabled, the text says to set MOSAEL_CLOUD_URL, and the cloud_not_configured machi…
- Fixes where the UI said one thing and did another. Section descriptions are no longer capped at 42rem with a big empty gap on the right; when FFmpeg is pinned by the environment the field's placeholder no longer contradicts the line above it; "recheck" moved i…
- Website: the sign-in entry is redesigned, and signup lands in the community. "Sign in" is a quiet outlined button next to "Download" instead of a bare text link, and a successful MosaelCloud signup now goes straight to the community instead of stopping on the…
Full notes and downloads
October 19, 2026
Stable
v1.9.10
- The agent is no longer cut off at 600 seconds. The timeout measures how long nothing has progressed instead of total runtime — text, thinking, tool start and finish, and every sub-agent step count as progress, so an agent that keeps working keeps going; a 2-ho…
- The inspector's task plan and background tasks share one row syntax. Status lives only in the circle on the left (done / in_progress / succeeded / queued all collapse into one state), actions sit on the right, and progress, elapsed time and failure reasons sta…
- What you picked is recorded in your own message. Once a choice card is answered the record belongs to the user message, leaving one noun-like line on the assistant side (“Picked …” / “Skipped”) instead of the assistant restating your choice in its own words.
- A stale translation engine id no longer 502s. The early ids (google / ai) are mapped to their canonical ids in one place; naming the AI engine or a translation plugin now really takes effect instead of being silently rewritten to Google, and an unknown id list…
- The notes top bar is quieter. The AI assistant is an icon now (the same entry as the editor's top bar), and “Saved” became one four-state icon — saved, saving, draft and error each with their own glyph and hint, readable by a screen reader — in a 30px slot ins…
- The full FFmpeg build list is refreshed. Windows moves to upstream autobuild-2026-10-19-13-06 while macOS / Linux stay on 9.0.2; this missed 1.9.9 and takes effect from this release.
Full notes and downloads
October 19, 2026
Stable
v1.9.9
- A big ComfyUI no longer hits “output too large”. Directory operations (tools, models, workflows, node packs, voices) got an 8 MB output budget instead of 1 MB, and the plugin side no longer silently trims a model list — it used to list only the first N; now an…
- Model downloads resume. Downloads support ranged resumes (Range / 206 / 200 / 416, reconciled with a sidecar file); cancelling or failing keeps the partial file, the next attempt continues from N, and you can explicitly delete the unfinished part; orphans unto…
- Remote connections show their version and open the workbench directly. The Open workbench entry no longer depends on the whole workflow library — it exists on the web build too — and it stays visible without an address while saying why it cannot be used. The c…
- Missing items tell the truth first. Twelve fixes: attribution from the graph's own cnr_id / aux_id / ver works offline; an output that “doesn't match” no longer poses as “nothing found”; declarations without a catalogue are not dropped; a version mismatch says…
- URL import gains three gates. A 2 GiB limit, a wall-clock timeout for the whole request, and sniffing the landed file header (an HTML page renamed to .mp4 is refused); plus a resolve() site-adapter seam and an image direct-link import entry; MissAV is adapted…
- Editor: no second audio track, no stuck progress. The lightbox no longer plays a second audio track from the library's neighbouring item; subtitle translation has a timeout and can be cancelled instead of sitting at 0/1039 forever; subtitle and transcript acti…
Full notes and downloads
October 17, 2026
Stable
v1.9.8
- The free Edge voice works again. Microsoft changed its token check on 2026-10-13 and older builds got a 403 on every request; this release follows the new rules, so Edge speaks again in the hands-free dock and in dubbing. Failures no longer dump a whole wss UR…
- Speech input can choose its engine. Settings → Agent → Spoken replies gains Speech input engine, defaulting to Follow transcription and listing only ready engines; if the chosen engine becomes unavailable you are told exactly what is missing instead of it sile…
- The hands-free dock shows it is working. From stopping the recording to the transcript arriving it shows Transcribing… with a spinner, so a first model load stays visible; closing the dock while waiting really cancels (the recorded sentence is not sent), and a…
- Spoken replies start sooner. Conversation speech is now a per-sentence pipeline: the first sentence plays as soon as it is synthesized — measured on real Edge here, first-sentence latency dropped from 3.285s to 1.567s (about 2.1x). Switch back to Wait for the…
Full notes and downloads
October 17, 2026
Stable
v1.9.7
- Feishu is finally dependable in this release. Bindings are per application (Feishu open_ids are scoped to the app, so with two bots in one workspace each person binds once per bot; creating a bot by QR code now binds its creator; bindings left behind by a repl…
- Messages and files in Feishu can start workflows. Drop a video into a group and a workflow runs and posts its result back; triggers are first-class data and a matching message is claimed outright (it is no longer also sent to the conversation, so no double spe…
- A new Local app settings page, and one grid for every asset field. Startup at login, opening external files, the version and update check move to Local app, leaving Backend with the server switch, API address and workspace id; the images a model should look at…
- Transcription gains a light engine and ComfyUI stops guessing. sherpa-onnx runs the same SenseVoice model without torch — about 9x faster cold and 32x warm (FunASR is still the default; sherpa-onnx has no speaker diarization) — and ComfyUI reads a workflow's o…
Full notes and downloads
October 14, 2026
Stable
v1.9.6
- The plugins page no longer crash-loops. Opening the plugins page (especially a ComfyUI connection) would crash the whole UI a few seconds later and keep crashing, and deleting then recreating the connection made it fine for a while. The cause: every refresh of…
- Editing exports choose a format, can export audio only, and no longer fail on an empty video track. The export dialog now asks video or audio first, then the format: MP4/MOV/WebM for video and MP3/WAV/M4A for audio, and the resolution, frame-rate and quality r…
Full notes and downloads
October 9, 2026
Stable
v1.9.5
- The workbench top bar is tidier. The canvas-input button now names the current mode (trackpad/mouse) and shares the outlined frame with Save and Run; dragging the right column's edge keeps the canvas edge exactly under your pointer with no extra blank strip; a…
- Scrollbars overlay on Windows and Linux too. Hover-to-reveal used to squeeze content narrower and back on those platforms (their native scrollbars take layout width); the desktop shell now enables Chromium's overlay scrollbars there, so behaviour matches macOS…
- Library detail pages look right. The preview in the left pane of model and workflow detail pages is now centred both ways instead of stuck to the top, and still scrolls normally from the top when the content is taller than the pane.
- Update checks say the real reason they fail. A dead proxy or a rate-limit collision no longer looks identical to “you're up to date”; failures are written to main.log, including the silent background checks. Eight more silent-failure paths now leave a trace (i…
Full notes and downloads
October 9, 2026
Stable
v1.9.4
- The design language is final, and the whole app follows it. Buttons, inputs and selects share one ruler; the on state has exactly one look (pressed states live in the base components — canvas toolbars, the editor and boards no longer paint their own); settings…
- The agent actually stops when you say stop. Queued messages no longer run on their own after you stop; which parts of the UI refresh after a card is approved is now declared by the tool itself; every creative session remembers where it was opened (boards, work…
- A string of ComfyUI fixes. Renaming or moving a workflow in its library updates everywhere Mosael uses it; old-format forms from older versions no longer silently run as the full workflow; a workflow's name is consistent everywhere; connecting through port for…
- Permissions and billing tightened. Read-only members no longer see write buttons everywhere only to be rejected on click; Clear finished only tidies your own task center instead of wiping the whole workspace's history, and spending remembers who paid; schedule…
- More fixes. An empty outbound-proxy field now tells the truth (follow the system proxy, local addresses always direct); workflow entries with forms in Add node each get their own selectable row; Civitai searches started by editors themselves show results; and…
Full notes and downloads
October 6, 2026
Stable
v1.9.3
- ComfyUI can live inside Mosael. If you already have one, pick its folder and a Python and Mosael starts and stops it, shows its logs, starts it when needed and stops it after a while idle to free the GPU; if you don't, choose Let Mosael install it, review the…
- Gemini can be the agent's chat model. Add an API key to the Google AI Studio connection and the agent, boards and workflow LLM nodes can use Gemini; Gemini 3's thought signatures are sent back on every tool call, thinking levels follow each model, and spending…
- ComfyUI works more smoothly. Running from the workbench changes the seed the way ComfyUI does; a Civitai lookup shows the source link and example image right away; when ComfyUI is on this computer, downloads go straight into its models folder with progress and…
- Exports no longer burn in "AI-generated" by default. The on-screen label is off unless you tick it in the export dialog for platforms that require it (top-left at the start, top-right throughout); the AIGC tag is still always written into the file. Failed gene…
- Product pitch template v7 and more accurate spending. The number of beats is capped in code by the target length, moving beats no longer pay for sound that is muted anyway, and the presenter's opening and closing lines get subtitles; cancelling a generation ca…
Full notes and downloads
October 5, 2026
Stable
v1.9.2
- ComfyUI workbench. The desktop app opens ComfyUI's own canvas with Mosael's model library, missing items, app form and run results docked beside it: select a loader node to list that folder's models and fill one in with a click; install missing node packs in o…
- App forms turn a workflow into a simple form. Pick the few inputs others should fill in, name and order them, narrow their choices, or start from a recommended set; AI Studio, boards and workflow nodes then show only those, and everything else runs as the work…
- The model library understands your models. Base models are recognized from metadata, weight structure and file names, including newer ones such as Krea 2, Anima and MiniMax H3; previews have four display levels with NSFW handled separately (manual marks, metad…
- Agent skills. Write how a kind of task is done as a SKILL.md (an open format) and the agent reads it only when needed; three built-in skills (creative boards, building workflows, 3D scenes) shrink the system prompt by about a fifth. Settings can create, import…
- Workflow folders and moving product beats. Workflows are grouped by ComfyUI's folders, dragging a card onto a folder moves it, you can create new ones, and cards have a context menu; with a video model, each product-pitch beat turns its frame into a short clip…
Full notes and downloads
October 5, 2026
Stable
v1.9.1
- Cloned voices can speak through Bailian CosyVoice. A voice from your library is cloned into your own Bailian account; with CosyVoice the voice dropdown gains a My cloned voices group for dubbing, subtitle dubbing, AI Studio, boards, workflows and spoken replie…
- CosyVoice voice-overs match their prices. Speech is recorded under the connection's vendor, so running Prefill prices once on the Bailian connection records CosyVoice at its list price; past records are fixed on upgrade, and older Edge voice-overs still marked…
- ComfyUI's image count becomes Runs, and long dropdowns stay quick. Each run makes what the workflow is saved to make (a saved batch of 4 gives 4), and the field says how many per run and in total; for workflows that only preview, only the final result comes ba…
Full notes and downloads
October 5, 2026
Stable
v1.9.0
- Notes get the AI assistant and a selection toolbar. Select a passage to polish, rewrite, expand, translate or continue it; changes to the text arrive as an original → new confirmation card and apply only once approved. You can also read aloud, highlight, quote…
- The browser session becomes a place to collect material. A row of icons in the top bar takes screenshots, downloads videos, collects images, saves the page as a note or starts work from it, and downloads you click on a page go straight to the library; an Arc-s…
- The ComfyUI plugin adds a model library and a workflow library. Browse the server's models (previews, base model, trigger words, who uses them) and workflows (node graph, inputs, what is missing), download missing models from HuggingFace, Civitai or ModelScope…
- Spending is recorded more accurately, and the media page is faster. Prices can be tiered by resolution, GPT Image is billed on the tokens the provider reports, Evolink on the charge in its response, failed calls that were never charged cost nothing, a workflow…
Full notes and downloads
October 3, 2026
Stable
v1.8.3
- The editor, rebuilt from the inside out. Dropping a clip overwrites what is under it, video stays linked to its audio, and ripple delete and speed changes carry what follows; several people or the agent can edit one timeline without overwriting each other, and…
- Subtitles, voice-over and export are sturdier and faster. Bilingual lines stack at the bottom, subtitle tracks import and export .srt / .vtt, text strokes no longer eat into white letters, and assets remember where they came from; exports stop clipping with op…
- The agent and boards feel smoother. Approval cards stay with the step that asked and fold into a line once decided, job receipts no longer crowd the message queue, and the agent no longer stops mid-sentence at a colon; on boards a ComfyUI workflow has a single…
- Workflows and analysis templates are more dependable. References show node and output names instead of raw braces, the start node and every property get their own ports, and a running node that references one that can never run is stopped before the run; Comme…
Full notes and downloads
October 2, 2026
Stable
v1.8.2
- Three analysis workflows: account health check, viral video breakdown and comment insights. Give a link and it gathers account and post data plus top comments, works out posting cadence, engagement, top posts and momentum, and writes a note; the breakdown can…
- Select several board cells and wire them from the + on the selection box, while each cell's own + connects only itself; structured plugin parameters such as Manim's explainer steps become one card per item instead of hand-written JSON.
- Another thorough pass: "Always allow in this session" now respects the risk tier, one pre-run check catches problems before any node spends money, board merges no longer drop what you just typed, undo no longer removes a running cell, and lip-sync reruns reall…
Full notes and downloads
September 30, 2026
Stable
v1.8.1
- A thorough pass over workflows, boards and the official templates. Every node, board cell and official workflow was reviewed: silent failures such as upstream text being re-interpolated, loop outputs rewritten on save, nested loops deadlocking and loops runnin…
- Select several board cells and wire them all at once, or drag out a new cell that takes them all as inputs — a set of reference images in one move; copy and paste cells with ⌘C / ⌘V; generation results no longer overwrite edits you are making.
- Confirmation cards are redesigned: what will happen, its consequences and the parameters are easy to read, code keeps its line breaks and highlighting, and the card stays with its result after approval. The agent's Edge voice-over and asset deletion work relia…
Full notes and downloads
September 30, 2026
Stable
v1.8.0
- Documents live in the library, and the agent reads them. Import PDF, Word, PowerPoint and Excel; they are parsed locally into page-by-page text with page images. Attach one in chat and the agent reads long ones section by section and looks at page images for l…
- An asset library of characters, scenes and props. Each named thing gets a wall of reference images by angle, with AI-filled angles and expressions; @ it when generating and its description and references come along; boards, workflows and the agent all know it.
- Digital humans: make a face talk, or re-sync a video's lips. Bailian, Kling, OmniHuman, HeyGen and Hedra side by side; templates for script-to-presenter, video translation with lip sync, and on-camera product pitches; real people need a consent declaration and…
- Plugins can register a capability anywhere. Denoise, vocal separation, transcription, translation, speech and document parsing are rows in one table where local engines and plugins sit side by side; the plugin market filters by capability; boards gain a timeli…
Full notes and downloads
September 27, 2026
Stable
v1.7.0
- Board cells do things themselves. Select an audio clip and transcribe, separate vocals and denoise sit right above it; videos turn into GIFs, notes and documents translate, and image-taking ComfyUI workflows live on image cells. Click one and the panel switche…
- 3D scene cells render their own blockout references: pick a shot and what to render, and the first/last frames and camera move land to the right. "Continue" is now a single "+ Generate" menu whose rows say exactly what happens.
- Picking a note, scene or asset uses one row-per-item dialog you can search and navigate with ↑↓ and Enter; generation pickers list only models that really make images or video, preselecting your default.
- "Ask AI to write" works again with models connected through account sign-in; ComfyUI names the model file when it lacks a text encoder; cell corners, run states and the generating placeholder are calmer.
Full notes and downloads
September 26, 2026
Stable
v1.6.0
- Music and sound effects in AI Studio. Write a description, put lyrics in their own field or switch to instrumental, and listen right in the player; Google Lyria, Suno, Kling, Bailian and Volcengine are wired in. Workflows, schedules and the agent can generate…
- ComfyUI is now a bundled plugin. Enter a server address and every saved workflow becomes both a model and a tool — parameters get readable names, reference images and masks are recognised, and every output lands in your library; existing connections move over…
- Boards can run tools that turn what you have into something new. Video to GIF, transcribe, translate, separate vocals, reduce noise, ComfyUI workflows and more sit on the board as tool cells, grouped by the content they work on; wire in notes, documents and me…
- A new Manim plugin turns formulas, graphs and code into explainer videos; object storage is one built-in plugin with S3-compatible support; the plugin market and workflow community become card grids with detail pages; the admin page splits into Overview / Memb…
Full notes and downloads
September 24, 2026
Stable
v1.5.2
- Follow what you trigger. A webhook trigger returns a run_id you can use with the same secret to check progress or cancel; reset the secret in one click if the URL leaks. Runs started from outside show up without reloading the page.
- An honest browser pool. A generic profile is a browser that keeps its state — Open returns to the page you left; publish accounts can really sign out, and Log in again signs out first. TikTok and WeChat Channels login detection is fixed.
- Hide objects and delete shots in 3D scenes; pulling from Blender brings its native cameras and lights, and scenes with imported models send to Blender again.
- One-click prices: 364 official list prices (chat, image, video, speech), each checked and cited, fill in whenever a provider's catalog has none. Agent memory is redesigned as a clear list, and the English interface now speaks English in errors, desktop menus a…
Full notes and downloads
September 24, 2026
Stable
v1.5.1
- Large screen recordings scrub smoothly. QuickTime .mov files are repackaged into .mp4 as they are imported (no re-encode, same quality), and existing ones are converted automatically; a single seek in a 370 MB recording went from up to 14.5 s to about 0.1 s, a…
- Clean camera recordings. Once access is granted the recorder shows a live camera preview and recording takes over that already-exposed stream — no dark opening, and mirrored recordings no longer freeze.
- Import in batches. Pick several files at once on the Media page and in the editor; filter by several tags; compare videos side by side, playing together and stepping frame by frame.
- Delete projects in bulk on Home, name a new project before it is created, see per-part titles when importing from a link, and get a clear reason (unsupported, nothing there, sign-in needed) when a link yields nothing.
Full notes and downloads
September 23, 2026
Stable
v1.5.0
- Mosael now records the post each publish created. The executor reads the post ID from the platform's own publish response (Douyin / TikTok, Bilibili, Xiaohongshu, YouTube) and stores it with the link and time: open it from the publish detail, wire the publish…
- Tencent Cloud COS official plugin: upload, time-limited links, pull back into the library, list — local assets become public links automatically. Signing uses COS's native q-sign, checked byte for byte against the official SDK.
- Remotion official plugin: animation from code. Lesson explainers, formula walk-throughs, code demos — give it the content and get a video, or write a component for anything freer, at no generation cost.
- Plugins can name their author and their guide: the plugin page and market show who wrote it, and Docs opens that plugin's own page.
- Regenerating a failed board node no longer looks like nothing happened: the job was submitted (and billed) while the canvas kept showing the previous failure, inviting a second click. One-click video from a 3D scene no longer copies the grey blockout — the pro…
- Local assets are linked through your own storage only: on shared deployments someone else's bucket and keys could be used before. With several set up, pick one under Settings → Asset links; the same asset reuses its link until it expires.
- Selected text copies again — the workflow canvas and the editor had taken ⌘C for nodes and clips. Home covers use the first frame on the timeline; Aliyun OSS signing and endpoints are fixed; dropdowns near the top of the window are clickable.
- Every dependency is on its latest release: Electron 44, Python 3.14, Node 24 LTS. The website's plugins and workflows are now community-style browse and detail pages, and downloads split into China (Baidu Netdisk) and global (GitHub).
Full notes and downloads
September 23, 2026
Stable
v1.4.3
- No new headline feature this time: it closes all 55 findings from five read-only audits. Those 55 share one trait — none of them raised an error: the code ran, the tests were green, the UI looked fine, and something was already quietly wrong.
- Conversations stuck on “thinking” forever are fixed: teardown cancelled the watchdog before waiting indefinitely on the child process, so a child that would not exit took the turn with it. Stop now really stops sub-agents, and steering/abort wait for the sidec…
- Local files are now uploaded to a public link automatically when used as a video reference. Some providers (e.g. Volcengine Seedance) only accept http(s) URLs for reference video; a local asset used to come back as an English 400. Three official storage plugin…
- Workflows no longer spend your money before telling you what is missing. A required field inside a loop body was invisible to the pre-run check, so “Topic to finished video” would run the brief, script, visual bible, storyboard and blockout (several paid calls…
- Fallbacks are no longer silent: dropping a tier because a model cannot do structured output or a thinking level used to fire in three places and be recorded in none. Max output tokens and thinking level are finally sent on directly connected models too — eight…
- Auto-approval leaves a visible trace: session settings list which actions this conversation auto-approved and which tier let each one through, and a mode set by someone else now says so instead of looking enabled.
- The home “AI usage” tile shows money, not call count, with a per-provider breakdown under the cost chart. The publish page now says when the desktop executor is not running — queued publishes simply never started, and nothing on screen said why.
- 3D scenes: the object list is a real tree (groups nest and collapse), a view-axis gizmo sits in the corner, the ground grid is shader-drawn, and groups fold with one click. Large Blender scenes no longer report a false “not responding” — the bridge had been bo…
- The browser extension speaks your language (it used to always ask the backend for Chinese). Deleting something still referenced is now consistently refused with the referrers named. Timeline operations went from 10 to 19 — the agent could previously use only h…
- The upgrade path was exercised for the first time, catching a migration ordering bug that crashed startup. After a restart, plugin invocation records and Blender transfers no longer sit at “in progress” forever, and the sidecar exits with the backend instead o…
Full notes and downloads
September 21, 2026
Stable
v1.4.2
- Blockout reference frames can finally draw your own models. A product, prop or machine modelled in Blender and brought into Mosael used to be visible only in the 3D workbench — in every composition reference the automated pipeline handed to the image and video…
- Imported 3D models now belong to the workspace, not to one scene: import a prop once and place it in any scene, from the new 模型库 group in the add menu. They used to hang off a scene — and a workflow creates a new scene on every run, so a Blender-built model co…
- New workflow node Available 3D props: it lists the workspace's models with their names and measured width/height/depth for the set designer, so an automatically built blockout can hold real props instead of boxes and cylinders. The dimensions are measured from…
- When the agent looks at a model in Blender it is no longer limited to six fixed angles: it can ask for any azimuth and elevation, move in, frame only the part it just built, and switch to see-through shading to check whether two parts actually intersect — at w…
- A failed turn no longer erases what that turn did. A conversation that ran for three minutes across a dozen tool calls, cut off at the last step, used to leave nothing but “the agent failed” — while those tool calls really happened. The record was never stored…
- An answered choice card stays a choice in the transcript: it used to be flattened into a sentence and posted as if you had typed it, losing the question and the options. The UI now shows what was asked and which option you picked.
- A running workflow can be stopped. The only place to cancel used to be the × on a row in the task centre: open the detail dialog to watch the progress bar and there was no way out, and the workflow page's run button stayed a ▶ while the run was in flight. The…
- Five new official templates: four business-facing ones (highlight shorts, product on model, product pitch, fabric lookbook) and the most common enterprise shape of all — montage from your own footage, with voice-over and subtitles, for when you already have th…
- Structured output is no longer a gamble: a json_object node without a schema used to 400 every time, and when a model returns malformed JSON the error is now explained before it gets one chance to fix it. Model settings gained a supports json_schema switch, pr…
- Fixed the infinite canvas going completely unclickable — React Flow's stylesheet went missing along with the lazily loaded page. Plus three other regressions from lazy loading, and two places that displayed “we don't know yet” as if it were a conclusion (the a…
- Sub-tasks now expand in place in the task detail dialog to show their progress, and opening a sub-workflow's canvas finally lights its nodes up as it runs.
Full notes and downloads
September 20, 2026
Stable
v1.4.1
- The agent can model inside your own Blender: read the scene graph, look at the viewport, run modelling code, and bring the result back into a Mosael 3D scene. It can see what it made — view_scene and images in tool results let it check its own work instead of…
- Code execution splits into two paths: sandboxed (compute something, process some text) and unisolated, on your machine (run_host_code — install things, touch files). They are two separate permission tiers, so opening up the first does not open up the second. E…
- Two new workflow nodes, “Build a 3D blockout” and “Render a blockout reference”, available to the agent as well; blockout rendering runs on the backend, so a workflow can use a 3D scene as a generation reference. The official “Topic to finished video” template…
- Generation nodes can pick which asset group to use, and a reference list can be connected as a whole instead of one wire at a time. Between nodes, a reference is a dependency: whatever a node reads through {{…}} is waited for before it starts.
- Paid remote generations are no longer cut short by a local timeout (ADR-0019): they end when the remote side actually finishes, or when you cancel. A failed item in a loop no longer starts a new paid iteration either — every failure is reported instead.
- Assets used by a timeline can now be deleted directly. The clips that use them stay where they are and become “media offline” placeholders: position, duration, transform and keyframes untouched, hatched red on the timeline, a red frame in the monitor, and expo…
- The agent can delete assets and projects too, on a new “permanent delete” permission tier: one batch at a time, with the card stating how many, which ones, and how many clips go offline — and this tier is never auto-approved.
- Model context and max-output limits are no longer guesses: a built-in table of verified limits (60+ entries across OpenAI, Anthropic, DeepSeek, Gemini, Grok, Kimi, GLM, Qwen, MiniMax, Doubao and common open weights) fills in whatever the provider catalog does…
- Fixed “the model used up this turn's output budget”: reasoning tokens share that budget, and a 1M-context model was being given only 16,384 — enough to be cut off mid-thought.
- Pages open much faster: routes are loaded on demand, taking the main bundle from 4.20 MB to 0.28 MB, and database migrations that have already run are recorded so startup no longer replays them.
- Fixed pricing rules being readable across workspaces — the read path did not go through the workspace check.
Full notes and downloads
September 18, 2026
Stable
v1.4.0
- New official workflow “Translated dubbing with subtitles”: transcribe and translate line by line, lay subtitles on the original timecodes, then dub each line and time-compress it back into its own slot. The original audio stays on its track, only ducked.
- The agent can dub subtitles from a conversation too. The confirmation card states how many subtitles are affected, whether the dub is compressed back to length, and what happens to the original audio — it is a paid action.
- What happens to the original audio after dubbing is now a choice of four: lower it while the dub speaks, mute it (for translated dubbing), leave it as is, or remove only the voice and keep the music. The editor, the agent and the workflow node all offer it, an…
- One run, one notification: a translated dub used to raise a dozen “done” toasts (one per spoken line, one for the proxy queued afterwards). Those now belong to the run that started them, the task centre announces it once, and proxy transcoding only speaks up w…
- New noise reduction: the built-in engine measures the noise before it acts and removes background hiss without touching music; for speech-led recordings, pick the open-source DeepFilterNet (best results, one download in Settings) or RNNoise (ships with the app…
- New voice / background separation: the media library, editor, workflow node and agent share one capability that produces two new assets and leaves the source untouched; the dubbing template uses it to mute only the voice and keep the music.
- Settings are regrouped by purpose; AI Studio gains an Audio tab for speech and podcasts; subtitle dubbing moves to the editor’s Voice tab and follows the timeline selection.
- 3D scene cards have thumbnails: a top-down plan drawn from the scene data itself, so it follows edits instead of drifting from a stale screenshot.
- Models the built-in catalog does not recognise can now carry parameters you write yourself: hand-typed aliases and the same model behind a relay used to show no parameters at all. Point each capability at a parameter set — and when a model is unrecognised the…
- Fixed dubbing failing to duck the original — two people talking at once in the export. When the source sits on a video track, ducking never applied (measured −0.03 dB; −10.35 dB after the fix).
Full notes and downloads
September 11, 2026
Stable
v1.3.1
- The agent composer supports @ references to assets, notes, boards and workflows: names in the text, ids in structured fields, and clickable chips after sending.
- Thinking levels now come from a per-vendor table: models that cannot stop thinking no longer offer "Off", and connections that cannot carry a level say why.
- Fixed menu separators being the same colour as the surface in dark mode — menus, the command palette and the notification and task panels all get their dividers back.
- The generation engine-settings panel is grouped into engine, output, inputs and tuning, with a header that no longer scrolls away.
Full notes and downloads
September 10, 2026
Stable
v1.3.0
- Long dropdowns now search: models, voices, fonts, LUTs, publishing platforms and upstream workflow outputs.
- Team discussions moved to a right-side sheet, shared by boards and workflows; workflows gained the discussion centre.
- Jumping to a comment or dropping a marker now avoids docked and floating side panels.
- Composer attachments and note references share one row, with thumbnails and click-to-preview.
Full notes and downloads
September 8, 2026
Stable
v1.2.0
- Mosael 1.2.0 正式版 / Stable release
- 统一工作台与完整物体时间线：相机和普通物体共用关键帧插入、拖动、多选、删除和撤销；移除重复的走位与记录视角入口。
- 自由视角、机位视角和全局动线同步预览；全屏保留建模助手、Blender、生成、导出及属性编辑。
- 场景列表保留详情导航，补齐右键与多选；右栏整体 section 可折叠，属性直接展开，统一单行时间线与面板间距，添加物体使用 ghost 按钮。
- 场景帧支持图片生成；视频可使用首尾帧或运镜预览。增加灯光预设、灰模参考、压缩模型支持及流式模型存储。
Full notes and downloads
September 7, 2026
Stable
v1.1.0
See the release notes for the full list of changes.
Full notes and downloads

Browse earlier releases