Changelog · Release history

What shipped, and when.

Every Nimbus8 release documented in one place. New modules, model support, runtime improvements, and bug fixes — newest first.

Format: reverse-chronological Also in public release notes

v1.2.7 — Nimbus lineup installs, thinking defaults, honest RAM July 2026

Nimbus 1.2.7 (iPhone / iPad) ships the Nimbus-Labs lineup polish that landed after 1.2.6. The matching Mac App Store build continues on the 1.2.5 version train with the same fixes.

  • Gale installs stay usable: Background downloads of Nimbus-Labs models register for the module you opened, keep vision tags, and no longer look “Not compatible with Gale” after they finish.
  • Hybrid thinking defaults: 2B and 4B stay quiet unless you opt into thinking; 9B keeps reasoning on by default, matching the Hub chat templates.
  • More honest RAM estimates: Catalog memory needs track the real MLX / GGUF formulas (4B ~4.4 GB class; 9B still needs a 16 GB+ machine — MacBook Neo stays on 4B).
  • Dictation and Cirrus polish: STT no longer re-glues the previous take into the composer, and Cirrus resume instructions stay out of the chat bubble.

v1.2.5 — Spire RSI clarity and local Diagnostics July 2026

Nimbus 1.2.5 strengthens local model-improvement workflows on iPhone, iPad, and Mac. Spire’s RSI runs are easier to set up and understand, completed runs give clearer next steps, and the new Diagnostics Trail makes local activity easier to inspect when you need it.

  • Clearer Spire RSI setup: The new-run flow gives more helpful, context-aware example cases and makes the relationship between the question, expected answer, and selected controls easier to understand.
  • More honest run progress and results: Spire now presents the rewriting and evaluation stages with clearer context, then reports before-and-after quality, judge availability, and the cases where a run still struggled.
  • Safer fine-tune handoff: Completed tuning artifacts are identified more clearly, helping you understand whether Nimbus created an adapter or a finalized local result and what can be used next.
  • Diagnostics Trail: A new, local-first activity trail records module and model lifecycle events, run stages, timing, and errors. Open it from Settings on iPhone and iPad or from the compact Diagnostics control on Mac.
  • Private by default: Diagnostics remain on your device unless you explicitly export them. You can search and filter the trail, clear displayed entries, or export a support-friendly text, PDF, or Word document when needed.
  • Better support visibility: Spire and runtime operations now add more structured local diagnostic detail, making it easier to understand what happened without exposing hidden model reasoning.
  • Localization coverage: New Spire and Diagnostics UI is included in Nimbus’s supported language catalog across iPhone, iPad, and Mac.
  • Release hardening: The iOS and macOS release archives now verify the production entitlements and Nimbus dSYM coverage required for a safer App Store submission.

v1.2.4 — Cirrus workspace polish, easier model switching, and steadier creative runs July 2026

Nimbus 1.2.4 is a focused reliability release for iPhone, iPad, and Mac. Cirrus feels more dependable in longer work sessions, model choices are easier to understand, and switching between local models is calmer across the app.

  • Cirrus is steadier: Workspace sessions, assistant responses, and model handoffs behave more consistently on iPhone and Mac, with fewer stuck or confusing states.
  • Model switching is clearer: You can stop a model while it is preparing, move to another module, and select an already-downloaded model from recommendations or search results without starting over.
  • Better memory housekeeping: Nimbus now gives model runtimes more room to finish shutting down before another one starts, helping prevent hangs when moving between Gale, Cirrus, and Mirage.
  • Mirage GGUF reliability: Stable Diffusion GGUF generation is more resilient to interrupted mounts and temporary memory pressure, with cleaner recovery for the next run.
  • Model browsing feels more honest: Installed, downloading, preparing, and ready-to-use states are shown more consistently across curated recommendations, Hugging Face results, and downloaded models.
  • Downloads and setup: Model installation, retry, cancellation, and runtime preparation now leave fewer stale progress indicators behind.
  • Mac parity: The Mac model picker and installed-model controls now follow the same safer switching behavior as iPhone and iPad.
  • Reliability throughout: This release includes additional fixes across Cirrus, model discovery, runtime cleanup, Mirage, and the shared iOS and Mac surfaces.

v1.2.3 — Grounded Gale answers, dependable downloads, runtime controls, and Nimbus Quick July 2026

Nimbus 1.2.3 is a broad reliability release for iPhone, iPad, and Mac. It makes everyday Nimbus interactions calmer and more dependable: better follow-up answers in Gale, clearer model setup, stronger downloads, more capable Mac controls, and a cleaner Nimbus Quick experience.

  • Gale follows the conversation better: Ask a follow-up question after a web-grounded answer and Gale now keeps more of the earlier context in mind. Search results are more focused, and the sources shown with an answer better reflect what Gale actually used.
  • Choosing a model is clearer: Nimbus does a better job of steering you toward models that fit your device and the module you opened. The model browser is faster, remembers choices more reliably, and gives clearer information before you download.
  • Downloads recover more gracefully: Large model installs are more resilient when interrupted. Progress, notifications, and the download activity view better reflect what is happening, and completed models are more reliably ready to use.
  • Nimbus Quick is more intentional: The compact Quick pill opens cleanly. After you type, results appear without taking over Return; pressing Return follows Nimbus’s normal routing, while the arrow keys let you deliberately pick a result.
  • Voice input is steadier: Nimbus is better at preparing and recovering speech models, so dictation and voice-driven tasks behave more consistently when a model has just been installed or needs to be made ready.
  • More polished creative and work modules: Cirrus improves workspace sessions and assistant output, Ashe improves task and skill handling, Mirage strengthens its generation queue, Mist improves downloads for offline knowledge, and Haze is easier to manage on Mac.
  • Better control on Mac: The Mac app gives you a clearer view of installed models and their readiness, making it easier to resolve an issue, switch models, or get back to work.
  • Reliability throughout: This release includes a wide set of fixes across model setup, downloads, sessions, generated content, and the main Nimbus surfaces on iPhone, iPad, and Mac.

v1.2.1 — Gale hotfix, Pro access, speech, Quick, and localization July 2026

Nimbus 1.2.1 brings the post-1.2 polish pass to iPhone, iPad, and Mac: safer Pro access recovery, a Gale reliability hotfix, better Apple Speech transcription, Mac Quick improvements, clearer local endpoint support, and a complete localization refresh across every supported language.

  • Gale reliability hotfix: Fixed the launch and inference paths that could leave Gale unable to answer correctly after the 1.2 update, including tighter fallback behavior around custom endpoints and local model routing.
  • Pro access recovery: Restore now refreshes access immediately, times out cleanly on bad networks, and recognizes existing lifetime unlocks instead of leaving paid users stuck on Free.
  • Apple Speech: Live dictation and one-shot transcription handle permission prompts, installed locale assets, finalized speech ranges, confidence metadata, and fallback recognition more reliably.
  • Mac Quick polish: Nimbus Quick received a broad Mac pass: cleaner panel behavior, sturdier insertion, improved settings surfaces, and better handoff from Quick into the full app.
  • LAN model endpoints: Custom OpenAI-compatible endpoints can now target private local-network hosts, .local names, link-local addresses, and Tailscale-style addresses while still blocking unsafe public cleartext HTTP.
  • Model and runtime UI: Model pickers, license tags, connector settings, MCP forms, Stratus, Mirage adapters, Overture, and browser rows received focused copy and state cleanup across iOS and Mac.
  • Localization refresh: Arabic, Bengali, German, Spanish, French, Hindi, Indonesian, Japanese, Portuguese (Brazil), Russian, Thai, Turkish, Urdu, Vietnamese, and Simplified Chinese were refreshed against the current string catalog.
  • Localization tooling: The Nimbus localization helper now audits encoding damage, placeholder parity, source-language fallbacks, invariant product/model names, and user-facing Swift literals more aggressively.
  • Release polish: The 1.2.1 builds received one more localization and stability pass across iPhone, iPad, and Mac.

v1.2 — Android and Windows foundations, runtime packs, and shared stores July 2026

Nimbus 1.2 laid down the first serious cross-platform foundation while keeping the Apple-native app moving: Android and Windows gained real module shells, runtime-pack awareness, localized resources, and shared Rust-backed stores for models, workspaces, Haze memory, and history.

  • Android app foundation: Added Android device profiling, module catalogs, model registry, Haze catalog, history catalog, Mist search catalog, Mirage catalog, workspace catalog, and a runtime-pack installer surface.
  • Android runtime packs: Introduced installable Android runtime-pack modules for llama.cpp, whisper.cpp, stable-diffusion.cpp, ExecuTorch, MLC, MNN, NCNN, and Kokoro ONNX so optional engines can be present only when needed.
  • Android UI build-out: Expanded the Compose app and home view model into a broader usable shell with module routing, model/runtime state, workspace state, and localized setup copy.
  • Android Stratus: Added a local Stratus analyzer path and tests so vision/task routing has a concrete Android-side implementation target.
  • Windows app foundation: Expanded the WinUI shell with real module surfaces, localized resources, and the first Windows-side flows for Nimbus modules and runtime routing.
  • Shared Rust stores: Added Rust-backed Haze memory, history, model registry, workspace, platform runtime, and runtime-pack APIs so non-Apple clients can share the same local-first data model.
  • Sidecar runtime routing: Updated the desktop sidecar RPC and supervisor surfaces for broader platform runtime discovery and management.
  • Cross-platform localization: Added Android string resources and Windows RESW resources for German, Spanish, French, Japanese, and Simplified Chinese alongside English.
  • Version alignment: iOS and Mac project versions were moved onto the 1.2 release train before the 1.2.1 hotfix pass.

v1.1.8 — Model downloads, Mirage recommendations, and Overture polish June 2026

Nimbus 1.1.8 brought a focused iOS release pass: safer model download progress and resume behavior, better Mirage recommendations on iPhone and iPad, faster model picker previews, cleaner Overture Converse controls, and refreshed Mirage graph and landing surfaces.

  • Release readiness: Nimbus is now prepared as the 1.1.8 iOS release, matching the current app marketing version.
  • Model download progress: iOS model downloads now rely on URLSession plus Nimbus Live Activities for visible progress, avoiding misleading system task-failed cards while downloads continue normally.
  • Download pause and resume: Paused downloads wait for real iOS resume data before resuming; if iOS cannot provide it, Nimbus shows a clearer Retry path instead of silently failing.
  • Mirage iOS recommendations: iPhone and iPad recommend smaller Apple Core ML palettized Stable Diffusion 2.1 and 1.5 models first, while SDXL options remain available through search.
  • Model picker previews: Inline Hugging Face previews are faster and better scoped by runtime filter, especially for Overture audio models and Mirage image models.
  • Overture Converse polish: Conversation scrolling is smoother, and tapping the voice orb can interrupt both thinking and speaking states.
  • Nimbus Quick web search: Quick now recognizes Bing and Reddit search providers and handles more natural provider phrases such as searching on, in, or via a named provider.
  • Gemma 4 output cleanup: Gemma 4 thinking-channel text is separated from the visible answer more reliably in MLX and chat output.
  • Model optimization guidance: Conversion planning now reports fuller disk and memory requirements before starting heavier Core ML or MLX optimization work.
  • Mirage graph and landing fixes: The bundled Mirage graph resources, CI verification, and landing CTA asset were refreshed so web and in-app graph surfaces stay in sync.

v1.1.7 — Mirage downloads, Ashe work queue, and voice June 2026

Nimbus 1.1.7 is the parity release for iOS and Mac: Mirage, Ashe, and Overture/Quick workflows were aligned with safer startup, clearer model choices, and stronger recovery. It also improves SDXL Turbo handling on iPhone, makes model downloads quieter and easier to retry, simplifies Ashe workflows, and adds editor and generation reliability improvements across Core ML, MLX, and sd.cpp image paths.

  • Release readiness: 1.1.7 release notes and packaging were finalized across web, service, and public release surfaces.
  • Improved SDXL Turbo support on iOS: Mirage now handles full SDXL Turbo Hugging Face downloads more carefully on iPhone and iPad, including the extra model files needed before the model is marked ready.
  • Better Mirage download recovery: If a large model install is interrupted, Retry can finish missing setup files without forcing a full redownload when the main model file is already on device.
  • Clearer model verification: Verify Install can repair missing model files when possible and shows a clearer failure when a download is incomplete instead of leaving Mirage stuck on "Preparing model."
  • Quieter model downloads: Starting or finishing a model download no longer plays an unnecessary alert sound, and returning to the app no longer creates duplicate download banners.
  • Smarter Mirage recommendations: iPhone and iPad now recommend the smaller Core ML Stable Diffusion 2.1 and 1.5 palettized models first, while SDXL options remain available in the model list. Mac keeps SDXL and the palettized models in its recommended Mirage set.
  • Mirage studio and editing: Added editable node-graph tools, theme-aware canvas controls, and improved Mac layout polish for the Mirage studio area.
  • Mirage model and runtime fixes: Improved Mirage routing and diagnostics for sd.cpp and Core ML image pipelines, including safe-mode params, runtime diagnostics, and clearer failure logging/ranking.
  • Core ML diffusion safety: Fixed routing and validation for Core ML models with sparse metadata, reduced eager probing, and made startup behavior safer when selecting Core ML image pipelines.
  • HiRes and cancel reliability: Fixed HiRes/image cancel and progress behavior on iOS so generation can be restarted consistently.
  • Diffusion startup recovery: Added launch-time diffusion recovery and serialization gates so model/runtime startup is less likely to hang in unstable paths.
  • Dictation and STT: Better startup loading states plus Apple Speech fallback behavior for Overture/Quick voice paths, with localized Apple Speech labels.
  • Ashe stability and cleanup: Ashe feels lighter and more predictable, with simplified work-queue flows, less clutter, and sturdier handoff behavior between iPhone and Mac.
  • Gale attachments: Photo attachments in Gale are now consistently available across Mac and iOS in more interaction paths.
  • Developer reliability: Fixed build/tooling and framework validation issues around MirageModelProfile, parakeet, and macOS LiteRT/CI stability.
  • Ashe calendar and reminders: Ashe can read and create calendar events and reminders when you ask, with permission, and all data remains on-device.
  • Overture dictation polish: Clearer "getting ready" states while speech models load; Apple Speech can step in when your mounted transcription model is not ready yet.
  • Echo and Nimbus Quick voice: Localized live preview labels and improved routing so Nimbus Quick picks the model you expect.
  • Mac Transcribe parity: The Mac Transcribe picker and model discovery now align with iPhone behavior, including the same core model-selection heuristics.
  • Mirage model fit: Better default sizes and capability hints when installing checkpoints from Hugging Face or Civitai.

v1.1.6 — Stratus routing fixes and Cirrus stability June 2026

Nimbus 1.1.6 brought focused Stratus and Cirrus reliability updates, stronger on-device chat quality for Gemma GGUF models, smarter prompt budgeting for small models, session titling fixes, and manual history rename on iOS and Mac.

  • Stratus routing: Fixed Stratus MLX VLM model routing so VLM-capable models land on the correct runtime and stop falling into unintended routes.
  • Model visibility: Corrected model-selection filtering and curated-model list behavior so the right Stratus-compatible models appear in picker flows.
  • Cirrus preview behavior: Fixed final Cirrus 1.1.5 behavior and improved preview stability across the flow.
  • Brand and workflow polish: Updated Cirrus preview branding surfaces and adjusted launch visuals/CTA treatment for cleaner landing transitions.
  • Module demo stability: Resolved blank iOS module-page rendering triggered by phone-only showcase toggles.
  • Gemma GGUF chat quality: Gemma 2, 3, and 4 now use the correct chat templates and sampling defaults so GGUF models stop looping broken role markers and answer normally.
  • Prompt budget for small models: Gale and Cirrus system prompts plus history windows clamp to tier-appropriate token budgets so tiny and small chat models are not overloaded.
  • Session auto-titling: Gale history titles no longer stay stuck on the first message — short prompts get AI titles after the first reply, and failed AI passes fall back to a readable date stamp.
  • History rename: Long-press a session on iPhone or right-click on Mac to edit the title yourself when auto-titling misses.
  • Release readiness: Refreshed model picker and landing handoff polish to reduce startup and setup friction between web and app workflows.

v1.1.5 — Nimbus Quick, menu bar actions, regional search, and hotkeys June 2026

Nimbus 1.1.5 introduced the latest Nimbus Quick, Mac menu bar, Gale routing, regional search, configurable hotkey, localization, Mist restore, and Mac model-picker enhancements. It keeps the new Quick workflow compact while making long answers, voice input, clipboard actions, regional provider discovery, and selected local models behave more predictably.

  • Nimbus Quick Gale routing: Quick now uses the Gale model you selected or last mounted for Gale conversations instead of falling back to Apple Foundation Models when another Gale model is ready.
  • Follow-up conversations: Quick keeps the compact conversation context needed for multi-step Gale follow-ups without leaving long invisible history on screen.
  • Markdown and long answers: Gale and connector answers in Nimbus Quick render as formatted markdown, stay scrollable, and handle longer responses without visually cutting off.
  • Visible voice transcription: Nimbus Quick now shows the live or final transcript while you speak, so dictation is no longer invisible while it is working.
  • Mac menu bar quick actions: Nimbus now has a menu bar item at the top of macOS for opening Nimbus Quick, starting dictation, summarizing the clipboard, translating the clipboard, asking about selected text, opening Nimbus modules, checking Ashe/Channels/run status, opening Settings, toggling Launch at Login, and quitting Nimbus8 without hunting for the main window.
  • Launch at Login for Quick: The Mac app can start in the background at login so Nimbus Quick, global hotkeys, clipboard monitoring, and the menu bar are ready immediately without forcing the main app window open.
  • Nimbus Quick appearance controls: Settings now let you keep the native macOS Liquid Glass pill, match the Nimbus theme, and choose the Quick pill size.
  • Locale-aware shopping search: Quick recommends regional retailers from the user's current macOS language and region while keeping every provider available globally by alias. Google, Amazon, and YouTube stay universal defaults; matched regions add providers such as Shopee and Lazada in Southeast Asia, Flipkart and Meesho in India, Rakuten/Yahoo! Shopping/Mercari in Japan, Coupang and Naver Shopping in Korea, Tokopedia in Indonesia, Tiki in Vietnam, Mercado Libre in Latin America, Trendyol in Turkey, Noon in Arabic-market regions, Allegro in Poland, and bol in the Netherlands/Belgium.
  • Provider icons: Search rows now show appropriately sized brand marks for Google, Amazon, YouTube, GitHub, Perplexity, Shopee, Lazada, Flipkart, Meesho, Rakuten, Yahoo! Shopping, Mercari, Coupang, Naver Shopping, Tokopedia, Tiki, Mercado Libre, Trendyol, Noon, Allegro, and bol on iOS and Mac.
  • Configurable hotkeys: Nimbus Quick opens with ⌥ Space by default, and the microphone starts with ⌥M. Both shortcuts can be changed in Settings → Nimbus Quick → Keyboard shortcuts, with duplicate and macOS registration warnings.
  • Dictation workflow controls: Quick settings now cover listen-on-open, dictation enablement, optional translation, source and output languages, focused-field insertion, Accessibility permission status, and fallback copying when insertion is unavailable.
  • Per-route Quick models: Quick model settings now expose Auto or explicit model choices for Mist search routes, Gale answers and rewrites, Ashe handoffs, and dictation when enabled.
  • Clipboard and selection shortcuts: Quick can summarize clipboard text, translate clipboard text, ask Gale about the current selection, paste or copy recent clipboard history, and clear clipboard history.
  • Localization coverage: Quick action titles, helper text, settings rows, command hints, provider recommendation text, dictation labels, and hotkey UI now have string-catalog coverage across every supported app language.
  • Dictation result cleanup: Quick clears stale result state after dictation finishes so the next voice run starts from the current transcript instead of old output.
  • Faster Mac model pickers: Module model selectors on Mac open and switch tabs more smoothly, including the Hugging Face filter tab.
  • Mist model restore: Mist now promotes the last selected model into the active mounted runtime on open, so the selector and actual search model stay in sync.

Nimbus Quick command list

  • Open and navigation: Open app, Open URL, Save Quicklink, Quick settings.
  • Web and shopping search: Search Google (g, google, web), Amazon (a, amazon), YouTube (yt, youtube, video), GitHub (gh, github), Perplexity (p, perplexity), plus regional providers by alias: Shopee, Lazada, Flipkart, Meesho, Rakuten, Yahoo! Shopping, Mercari, Coupang, Naver Shopping, Tokopedia, Tiki, Mercado Libre, Trendyol, Noon, Allegro, and bol.
  • Local content: Search files, Add folders, Search snippets, Save snippet, Search Notes, Create Note, Append to Note, Delete Note, Search clipboard, Clear clipboard history.
  • Writing and AI: Improve Writing, Fix Spelling and Grammar, Explain Simply, Professional Tone, Friendly Tone, Find Bugs in Code, Save AI Command, Write with Gale, Search documents with Mist, Run Ashe, Run Cirrus.
  • Voice and output: Start dictation, Read Selection Aloud, Stop Speaking, Insert result, Copy result.
  • Calculator and Mac controls: Calculator, Lock Screen, Sleep Mac, Show Screen Saver, Sleep Displays, Restart, Shut Down, Log Out, Toggle Mute, Volume Up, Volume Down, Open Trash, Empty Trash, Eject All Disks, Hide Other Apps, Unhide All Apps, Toggle Appearance.
  • Window controls: Maximize Window, Left Half Window, Right Half Window, Top Half Window, Bottom Half Window, Center Window, First Third Window, Center Third Window, Last Third Window, Top Left Quarter Window, Top Right Quarter Window, Bottom Left Quarter Window, Bottom Right Quarter Window.
  • Menu bar actions: Open Nimbus Quick, Start Dictation, Summarize Clipboard, Translate Clipboard, Ask About Selection, Open Nimbus, New Chat in Gale, Open Ashe, Open Cirrus Workspace, Open Models, Open Channels Inbox, Settings, Launch at Login, and Quit Nimbus8.

v1.1.4 — Nimbus Quick and smarter local search June 2026

Nimbus 1.1.4 brings Nimbus Quick to Mac — a floating launcher on ⌥ Space — plus snappier Mist search and a round of polish across model browsing, fine-tuning, and Shortcuts on iPhone, iPad, and Mac.

  • Nimbus Quick (Mac): Press ⌥ Space from any app to search files, open apps, run saved snippets and notes, calculate, dictate, ask Gale, search Mist, or hand work to Ashe and Cirrus — all from one small panel.
  • Quick settings: Pick search folders, pin a mode, assign models per workflow, optionally start listening when Quick opens, and use Quick in 16 app languages. Voice, insertion, and full module handoffs are part of Nimbus8 Pro.
  • Faster Mist search: Results show up sooner while your library keeps indexing in the background; offline packs are easier to browse and download; search recovers cleanly if an index needs to rebuild.
  • Clearer model browsing: Hugging Face rows show better storage estimates and steer audio and multimodal models toward the right modules.
  • Spire fine-tuning: Smoother dataset import and training on iPhone, iPad, and Mac.
  • Stratus: More dependable on-device classification for vision workflows.
  • Ashe shortcuts: Trigger common Ashe actions from the Shortcuts app.

Read the full write-up: Option+Space, then speak.

v1.1.3 — Overture Converse, voice catalogs, and model routing June 2026

Nimbus 1.1.3 is the Overture and model-runtime sync on top of 1.1.2. Voice workflows are easier to pick, provider-managed audio models route more cleanly, and model browsers now make better use of capability and provider metadata before offering a model to a module.

  • Overture Converse: New conversation engine and voice picker surfaces on iPhone, iPad, and Mac, with shared voice-orb behavior across Overture and Ashe voice flows.
  • Provider-managed audio engines: Overture routes STT, TTS, audio analysis, audio generation, and speech-to-speech engines more reliably across FluidAudio, MLX audio, Parakeet, Kokoro, OpenRouter, HF Inference, and fal.ai paths.
  • Voice catalogs and imports: Added a Converse voice catalog, OpenRouter voice import support, expanded Kokoro voice metadata, and cleaner voice selection in Overture and Chinook.
  • Model picker parity: Module pickers now share stricter selection policies, clearer runtime labels, installed-state chips, and capability checks across iOS and Mac.
  • Better model discovery: Hugging Face browsing uses provider-inference metadata and repo-size fallbacks so audio and multimodal rows appear in the right module pickers.
  • Mirage and Mac sync: Mirage on iPhone, iPad, and Mac now pick models and runtimes the same way as the rest of the app.

v1.1.2 — Mirage reliability, Cirrus sandboxing, and runtime routing June 2026

Nimbus 1.1.2 carried the full Apple-native 1.1 reliability pass: Mirage image and video generation became more dependable across sd.cpp, MLX, and Core ML paths; model browsing hid formats your device cannot actually run; Cirrus Remote became easier to pair and control; and Mac sandbox support was tightened for local workspaces.

  • Mac Cirrus: Sandboxed Mac builds support app-managed local workspaces, security-scoped user folders, allow-listed command tools, terminal sessions, and safer workspace cleanup.
  • Mirage generation reliability: Lazy sd.cpp mounting defers multi-gigabyte weight loads until first generate; sd.cpp context invalidates when ControlNet, LoRA, or VAE selections change; SD 1.5 single-file checkpoints route cleanly without stalling on Diffusers folder installs.
  • OpenRouter media: Cloud image and video generation retry on transient failures; Fusion and managed-router models skip incompatible API fallbacks; dedicated media client extracted from Mirage view model for clearer routing.
  • Core ML Mirage: ControlNet wiring for the Apple ml-stable-diffusion path; img2img UI gated to runtimes that support it; auto-switch between image and video mode when the mounted model changes (iOS and Mac).
  • Smarter model browsing: Hugging Face and Civitai lists hide models Nimbus cannot run on Apple Silicon by default — unsupported formats, needs-conversion repos, and warmup-only runtimes stay out of the default list. Enable Advanced in Settings to reveal them.
  • Runtime switching hardening: Chat fails fast with a clear message when a mounted model has no on-device inference consumer; mount refuses unsupported and needs-conversion runtimes instead of marking them ready; inference mode routes by resolved runtime, not filename guesswork.
  • Cirrus Remote: Expanded remote session control — safer pairing, clearer run/test controls, better session recovery, and more predictable command consent when driving Cirrus from another device.
  • Download and install recovery: Mac disk pre-check before large installs; registry install pipeline consolidated across iOS and Mac; migration v2 for mounted-model persistence; auto-resume checks cellular policy and resume blobs.
  • Overture warmup syncing: FluidAudio provider preparation integrated into mount warmup so Overture cold boots no longer appear stuck.
  • Hardened Ashe approvals: One-shot SQLite-backed approvals prevent duplicate tool executions; centralized approval routing resolves sheet collisions; sensitive fields are automatically redacted; "Decide Later" snoozes prompts for 15 minutes.
  • Browser and accessibility polish: "Use Model" verifies variant compatibility before mount; model tiles support long-press quick actions; misleading "Dismount" labels replaced with accurate installed states.

v1.0.9 — Cirrus Remote, audio, and model polish June 2026

Nimbus 1.0.9 is a release-readiness update for the Apple-native workflow. Cirrus is easier to control across iPhone and Mac, model installs are checked more carefully before use, and the audio, image, and project flows received a broad stability pass.

  • Cirrus Remote control: Pair and manage remote Cirrus sessions with clearer QR setup, safer command pairing, better session recovery, and more predictable run/test controls.
  • Project workflow polish: Project state, deployment settings, and workspace handoff are more resilient, especially when moving between iOS, Mac, and connected cloud build or deploy targets.
  • Better model discovery: Hugging Face and Civitai browsing now sort and refresh more visibly, filter out more models Nimbus cannot run, and show clearer readiness states before install or launch.
  • Safer model verification: Nimbus performs stronger checks on mounted model bundles, Mirage image pipelines, speech-to-text models, and fallback routes so broken or incompatible downloads fail earlier with clearer errors.
  • Overture audio improvements: Dictation, speech, audio cleanup, and remote audio routes have better readiness checks, localized status text, and more reliable provider routing.
  • Mirage generation fixes: Image generation now cancels more cleanly, avoids competing Core ML pipeline ownership, validates transformer bundles more carefully, and releases audio resources after generation work.
  • More localization coverage: The app translation framework now covers 15 app languages, with more localized error and readiness messages across model, audio, and remote-session screens.
  • Everyday Mac and iOS fixes: Native copy actions, chat send timing, Haze memory writes, Live Activity progress, Mac command consent, simulator scaling, and release packaging all received polish.

v1.0.8 — Model downloads, Mirage reliability June 2026

This release focuses on making model installs feel calmer and more dependable, especially for Mirage image models. Nimbus now does more checking before a download is marked ready, shows clearer model choices, and avoids offering files the current app cannot actually open.

  • More reliable installs: Nimbus checks that a model has the files it needs before calling the download complete. If something is missing, the app fails early with a clearer message instead of letting the model break later.
  • Better Mirage support: Mirage can recognize and prepare more image-model layouts from Hugging Face and Civitai, including full pipeline folders and multi-file image models.
  • Cleaner model browsing: Model rows now do a better job showing the right file, size, and provider branding, with fewer unsupported choices mixed into the install list.
  • Smarter image runtime selection: Nimbus can route supported Mirage downloads through the right local image engine, including the newer MLX image path where available.
  • Download progress polish: Downloads and generation Live Activities now use Apple's system styling, so they fit better on the Lock Screen and Dynamic Island.
  • Mac and iOS fixes: Model picker, Mirage, settings, project, and download views received a round of small stability and layout fixes across both Apple platforms.

v0.14 — Stratus beta, Gale vision April 2026

Stratus enters public beta with camera-to-ask OCR powered by Apple Vision and on-device VLMs. Gale gains multimodal input — attach photos and PDFs alongside text prompts, with vision routing handled automatically by the capability manifest.

  • Stratus (Vision): Live camera OCR, photo library import, clipboard image paste. Routes through Apple Vision framework first, falls back to on-device VLM for complex layouts.
  • Gale vision: Attach up to 6 images or PDFs per message. The capability manifest checks whether the active model supports vision and routes accordingly — no manual toggle.
  • Thinking blocks: Models that emit <think> reasoning (DeepSeek R1, Qwen3-Thinking, QwQ) now render as collapsible disclosure panels in the chat bubble. Thinking is stripped from memory and context to save budget.
  • Streaming think filter: New incremental parser handles <think> tags split across SSE chunks without visual glitches.
  • Model registry: Persistent index.json tracks every installed model's repo, filename, module, SHA256, and last-loaded timestamp. TOFU verification on first load; mismatch on subsequent loads is a hard error.
  • Bug fixes: Fixed workspace engine state loss across FFI calls. Fixed streaming chat returning zero token counts. Fixed telemetry buffer growing without bound.

v0.13 — Mist knowledge packs March 2026

Mist gains offline knowledge packs — installable content bundles (Wikipedia, Wiktionary, Wikivoyage) that feed the hybrid search index. Download once over WiFi, search forever without a connection.

  • Pack catalog: Bundled JSON manifest with download URLs pointing at Wikimedia's own servers. Nimbus8 never rehosts pack content.
  • Streaming ingest: Tar+gzip archives are parsed and embedded in a streaming pipeline — parser on a blocking thread, embedder on the async pool, bounded channel for backpressure. A full Simple Wikipedia (~200k articles) indexes in under 10 minutes on iPhone 15 Pro.
  • Source filtering: Hybrid queries accept include_source_prefix / exclude_source_prefix so the UI can run "only packs" and "only my notes" lanes against the same index.
  • Uninstall: One prefix scan removes every document indexed under a pack. Idempotent.

v0.12 — Chinook TTS, Ashe scheduler February 2026

On-device text-to-speech via Kokoro-82M lands as the Chinook module. Ashe gains a production scheduler with exponential backoff and iOS background task integration.

  • Chinook (TTS): 54 curated voices across 10 languages. Kokoro-82M runs through ONNX Runtime on the Neural Engine. Output is 24 kHz mono WAV — playable directly through AVAudioPlayer.
  • Ashe scheduler: Per-hand interval ticking with exponential backoff on failure (cap: 6 hours). Foreground loop via tokio, one-shot tick_all_due for BGAppRefreshTask.
  • Fuel metering: Token budgets, wall-time limits, and tool-call caps per turn. Background tasks get tighter budgets than foreground interactive turns.
  • Widget snapshots: Ashe writes atomic JSON snapshots to the App Group container after every wake. The widget extension reads them without opening the full app.

v0.11 — Mirage img2img, Echo word timestamps January 2026

  • Mirage editing: Use any generated image as a starting latent. Describe the edit in natural language ("make it sunset"), adjust strength, and the pipeline runs an img2img pass.
  • Echo word timestamps: Per-token timing data from whisper. Enables word-by-word playback highlighting and SRT/VTT export with word-level granularity.
  • Transcript export: Export dictation sessions as SRT, VTT, or plain text from the Echo module.
  • Gallery multi-select: Long-press to enter selection mode, drag across cells to extend. Share, save to Photos, or delete in bulk.

v0.10 — Cirrus GitHub PRs December 2025

  • GitHub Device Flow OAuth: Authenticate with GitHub from the app using the device code flow — no web view, no redirect URI.
  • PR creation: Cirrus can open pull requests against your repos. Diffs are generated by the Rust diff engine (Myers algorithm), reviewed in-app, and pushed with your explicit confirmation.
  • Syntax highlighting: Code blocks in chat and Cirrus use Highlightr when linked, with automatic language detection.

v0.9 — Haze memory, model registry November 2025

  • Haze (Memory): Cross-session vector recall backed by SQLite FTS5. The turn loop injects relevant memories into the system prompt so the agent has continuity across sessions.
  • Hot memory: Key-value working memory (Hermes MEMORY.md pattern) with automatic extraction every N turns. Safety-scanned before persistence.
  • Skill distillation: After a successful turn, the model optionally distills a reusable skill from the transcript. Skills are FTS-indexed and injected into future turns when relevant.

v0.8 — Zephyr translation October 2025

  • Zephyr: On-device multilingual translation. Pick a source and target language, paste or type text, get a translation from the active model. No network required.
  • Adaptive engine: Thermal monitoring adjusts inference parameters (batch size, thread count) when the device heats up, preventing thermal throttling from degrading the experience.

v0.7 — Overture dictation September 2025

  • Overture (Echo): On-device speech-to-text via whisper.cpp. Supports WAV file input, raw samples, and base64-encoded audio. Auto-detects language or accepts a hint.
  • Audio capture pipeline: AVAudioEngine tap resamples hardware audio to 16 kHz mono in real time. Smoothed audio level published for waveform UI.

v0.6 — Mist hybrid search August 2025

  • Mist: On-device semantic search with hybrid BM25 + dense vector retrieval fused via Reciprocal Rank Fusion. Backed by tantivy (BM25) and ONNX Runtime (BGE/MiniLM embeddings).
  • Web search reranking: Fetch snippets from DuckDuckGo/SearXNG, embed them locally, and rerank by cosine similarity against the query. No search history leaves the device.

v0.5 — Mirage on-device diffusion July 2025

  • Mirage: On-device image generation via Apple's ml-stable-diffusion. SD 1.5 and SDXL Turbo on the Neural Engine. Private gallery with metadata, share, and save to Photos.
  • Memory guard: Refuses to start generation when available memory is below 400 MB. Prevents OOM kills mid-generation.

v0.4 — Ashe agent runtime June 2025

  • Ashe: On-device agent runtime with a turn loop (plan → tool calls → apply → reflect), audit log with Merkle hash chain, and a tool registry with per-tool access policies.
  • Loop guard: Detects repeated identical tool calls and breaks the loop before fuel is wasted.
  • Safety scanner: Regex sweep for role-hijack prefixes, invisible Unicode, credential exfil patterns, and SSRF probes in any text injected into the system prompt.

v0.3 — Cirrus code module May 2025

  • Cirrus: Code-focused chat with inline diff cards, file creation, and project-aware context. Diffs use the Myers algorithm via the Rust diff engine.
  • Workspace engine: Local project + file state on disk. Create, edit, delete files; build a file tree for the model's context window.

v0.2 — HuggingFace browser April 2025

  • HF discovery: Search, browse, and download GGUF and MLX models from Hugging Face. Variant picker shows quantization, size, and backend compatibility.
  • Resumable downloads: Backed by hf-hub with HTTP Range resume and LFS redirect handling. Compatible with the Python huggingface_hub cache layout.
  • Gated model support: Paste your HF token in Settings to access gated repos (Llama, Gemma). Token stored in iOS Keychain, never logged.

v0.1 — Gale chat March 2025

  • Gale: The first module. Streaming chat with any open model via MLX on Apple Silicon. Token-by-token output, conversation history, session persistence.
  • Rust core: Foundation library with FFI bridge to Swift. Inference client, diff engine, download manager, model manager, telemetry buffer.
  • Vanilla Wood theme: The default (and only) palette. Warm paper tones, quiet surfaces, no dark mode yet.