Singularity Node
Release notes
- New (beta, Mac): make short videos on your Mac in Create. Fast model by default; bigger models appear on Macs with more memory. Runs locally.
- Fixed: coding agents such as OpenCode now work with local models. Tool calls come back as real tool calls instead of raw text, so files get read and commands get run. Works on MLX and llama.cpp.
- Fixed: a local model could repeat the same tool call over and over. Repeated identical calls are now stopped.
- New: every model has a short ID (for example qwen3-coder-30b-a3b), shown on its card with a copy button. Old long IDs still work.
- New models: Qwen3.6 35B A3B and 27B, Qwen3.5 9B, Gemma 4, gpt-oss 20B and 120B, Nemotron 3.5, Granite 4.2, Ministral 3, GLM-4.7 Flash and more, on MLX and llama.cpp.
- Fixed: Gemma 4 12B downloads, and wrong sizes for several large models. Older models are hidden by default; choose "Include older" to see them.
- Fixed: "Start chatting with Apple Intelligence" no longer flashes the model list first.
- Fixed: the fly could get stuck circling in a figure 8 near a screen corner while Break enforcer was fighting your pointer. It now always lands.
- New (beta, Mac): your fly. A house fly that lives on your screen, driven by a real fruit fly brain (the FlyWire connectome, 138,639 neurons) running on your Mac. Open it from the fly icon at the top. Free, and it runs only on your Mac.
- Give it a job: Desktop cleaner, Downloads janitor, Screenshot sorter, Focus buddy, Break enforcer, Big-file sniffer, Duplicate finder, Battery and heat, Grid companion, or keep it as a pet. Each job lists exactly what it may touch, and the app enforces it. Anything it eats goes to the Trash, so Put Back always works, and nothing moves unless you switch that on.
- Teach it: turn on Training in its Settings, then give it a cookie right after it does something you like, or shoo it off when you do not. It learns your taste in files and its favourite spots, and gets tamer.
- Let it hear you (optional): shoo, blow or clap and it flies off. Set it up with your own sounds in a short guided step. Only the shape of a sound is used, live; nothing is recorded or sent.
- Feed it cookies (Command + Option + T, then click), watch its brain light up live, and pick its look.
- Log in with wallet from the account menu when you use Local without an account.
- Fixes: the model menu no longer fades its Download buttons; more space above Generate embedding.
- Linux beta: the app now runs on Linux (Ubuntu and Debian .deb, or an AppImage for other distributions). Chat and the Grid work; voice, dictation and image tools are Mac only for now.
- Windows catches up: the Windows app is now on this version, with everything from the releases since 1.10.5 that applies to Windows. Voice, dictation and image tools are still Mac only.
- Fixed a sign-in bug that could leave your node unable to serve. Starting a second sign-in while the first was waiting could replace your node's key, so the Grid rejected everything the node sent and paused it. The key is now saved only after the Grid accepts it, and a second sign-in is refused while one is open.
- If the Grid does not recognise your node's key, the app now says so and offers Log out and reconnect. Your stake is not affected.
- The "Couldn't go live" message now goes away on its own once your node is serving, and when you try again.
- Set your own price for the pictures your Mac serves on the Grid, from half to double the suggested rate. You see the price in dollars and you keep 80%. The editor appears in Pricing once price setting is switched on for the network.
- Image and text to speech models can now sign in to the Grid. Before, choosing one on the Grid setup step failed with "Unknown model" before the browser opened.
- Clearer sign-in errors: the note about 50,000 $SGL staked now shows only when the wallet approval itself failed.
- Grid dictation accepts a node's own price within the same range, and you can raise your per-recording limit up to $0.012.
- Image & Video: a new tab to make images on your own Mac. Pick a model, type what you want, edit any picture with words, upscale it up to 8K, remove the background, and save as PNG, JPG, HEIC or transparent PNG. Everything stays on your Mac. Video is next.
- Drag a picture from Finder onto the window, or paste one, to edit it. Edits from the Library continue the chat the picture came from, or start a new one.
- Library: search, filter, mark favourites with a star, select several pictures to delete at once.
- Image chats are saved in Recents with an image icon and reopen where you left off.
- Pin chats and sort them into your own groups. Drag chats to reorder them or move them between groups. Works for text chats and image chats.
- Pictures in chat: attach, paste or drop a picture for vision models.
- Switching the Local AI model now runs in the background, so you can keep using the app while it loads.
- Any Mac that can run the image model can serve pictures on the Grid and earn. The default is about $0.0063 for a 1024 picture, and you keep 80%. Windows and Linux with NVIDIA GPUs are coming soon.
- More voice models: MOSS-TTS Nano, LongCat-AudioDiT 1B, VoxCPM2 2B, and Rumik 3B for Hindi and 21 more Indian languages (personal use only).
- Fixed: Soprano 80M now shows up, closing the window and clicking the Dock icon brings it back, and model card Download buttons are readable on hover in every theme.
- Voice Studio: type anything and hear it in a natural voice, made on your Mac. Set a style or describe the feeling, and add audio tags like [laugh] and [sigh].
- Clone a voice from a 5 to 10 second clip, or design a new one from a sentence. Your recordings never leave your Mac.
- Up to ten speakers in one take, each with their own voice, model and style.
- Read aloud anywhere: select text in any app and press Control Option S.
- Ten voice engines, including Qwen3-TTS, Chatterbox, Orpheus, Dia, Kokoro, Phonon and the new Soprano 80M.
- Serve Kokoro text to speech and Whisper Large Turbo transcription on the Grid and earn per request.
- Local dictation now runs up to 10 minutes per recording.
- The running model has a green border on every model tab, voice previews play on hover with a ring that follows the audio, and the Grid sidebar shows only Grid tabs.
- Windows stays on 1.10.5 for now.
- Apple Intelligence is built in. On Macs with Apple Intelligence (macOS 26 or later), chat with Apple's on-device model in Local mode. Nothing to download, and nothing leaves your Mac.
- A fresh design: a calmer window with a floating content panel, a serif welcome, and a heartbeat that shows when your model is working.
- Colour themes in Settings: Default, Mono, Forest, Sand, Ocean, Plum and Rose, each in light and dark.
- Insights, redesigned: big-number tiles, a daily activity chart, model usage, a streak calendar, and a Your voice tab for Dictation.
- Four new share cards drawn from your own numbers.
- A Get started checklist for Local mode, a Help menu with What's new, and Settings in one window that checks for updates on its own.
- Collapsible sidebar, a compact Dictation pill, clearer Grid setup, and Open chat straight from the status card.
- Fixes for the model menu, narrow windows, links that did not open, and voice models briefly showing Download at launch.
- Windows stays on 1.10.5 for now.
- Run and serve Singularity xMESH 0.8B, our x402 routing decision model, from the Models tab.
- On Apple silicon, Run Locally places xMESH on the GPU. Serving on Grid uses the GPU with node v1.10.4 or later, which picks the verified path automatically.
- Fixes Serve on Grid and Run Locally for xMESH, which did not start in 1.10.6.
- Small spacing fixes on model cards and the switch-model dialog.
- Windows stays on 1.10.5 for now.
- This build was withdrawn from the public updater. Version 1.10.7 replaces it.
- Added Singularity xMESH 0.8B to the Models tab, but Serve on Grid and Run Locally failed to start it.
- Models now sits directly below Overview, with clearer navigation and separate icons for Pricing, Earnings, Jobs, and Insights.
- On Mac, private Dictation stays ready while a different Local AI model and a Grid model run at the same time. Each mode remembers its own model choice.
- Existing Mac Grid services may need one Reconnect Grid action to move to a separate port before Local AI starts beside them. The app shows this action when needed.
- Right Option or a recorded shortcut always uses the selected local voice engine, even while Grid serves. Mac Dictation shows live words and keeps a local history with copy and phrase corrections.
- Mac operators can download and serve the verified Whisper Small voice model on Grid. Pricing shows its fixed per-audio-second rate and the operator share.
- The compact recording indicator has a wider, evenly lit waveform, and Local Overview keeps Start, Restart, and Stop together.
- Windows receives the shared Node workspace improvements. Dictation remains disabled on Windows pending hardware validation.
- This build was withdrawn from the public updater. Version 1.10.5 replaces it.
- The app now opens directly to the established Node console, with Local AI, Grid mode, models, jobs, earnings, and settings in their familiar places.
- Dictation remains available as an optional sidebar tab and keeps the same local and encrypted Grid transcription support introduced in 1.10.1.
- No model, node, billing, privacy, or Windows feature gate changed in this interface-only hotfix.
- This build was withdrawn from the public updater. Version 1.10.5 replaces it.
- The Mac app opens an independent Dictation workspace with Local Whisper Small and encrypted Grid transcription modes.
- After the verified model download, Local dictation works offline without a wallet, login, API key, Grid provider, or node setup.
- A supervised worker isolates the local speech engine from the interface. Cancellation, timeouts, crashes, and app exit stop and reap the owned worker.
- Grid dictation uses a session-only consumer API key and explicit rate, privacy, and per-utterance consent. Audio travels encrypted to the selected provider.
- The provider processes the audio and returns a signed encrypted result. Uncertain payment outcomes require reconciliation before another paid utterance.
- Mac field insertion checks the original editable target. Secure fields, terminal targets, and changed focus remain copy-only or refused.
- The installed app and live encrypted Grid path passed release verification. No speed or accuracy rating is claimed.
- Windows dictation remains disabled pending real hardware validation. Public builds keep private canary-token controls disabled.
- Apple Silicon Macs can run EmbeddingGemma 2 locally or serve it to the Grid for text, image, audio, video, and ordered mixed-input embeddings.
- The new Embeddings workspace supports retrieval query/document modes and normalized 768, 512, 256, or 128-dimensional vectors.
- EmbeddingGemma 2 uses immutable verified model and runtime revisions, bounded media decoding, private temporary files, and readiness checks across every supported modality.
- Cached model verification now runs without freezing the desktop interface, so model pickers and Local AI startup remain responsive during integrity checks.
- Grid Vision is shown as an available shipped capability instead of “coming soon”; Qwen vision model cards still require the verified live catalogue and projector checks.
- Grid service setup keeps Local and Grid runtimes isolated and persists the dedicated embedding worker across app restarts.
- Windows receives the shared startup, download, and lifecycle reliability improvements; EmbeddingGemma 2 remains available only on supported Apple Silicon Macs.
- This app release pairs with SGL Node v1.9.17 for encrypted multimodal Grid routing and exact input-only usage accounting.
- Local AI startup now cleans up stale local sidecars before switching between Grid, MLX, CLM, and System One runtimes.
- CLM/System One cache checks now accept valid installed runtime marker files instead of asking users to download again.
- Model downloads now have separate Pause and Cancel controls: Pause keeps resumable progress, while Cancel clears partial files or MLX/CLM cache state.
- Download cards clear stale progress after completion or cancellation, so users do not see phantom 95% download states.
- This Mac release replaces the previous Local AI build. Windows remains on the current stable release.
- Run Laya/System One directly in Local Mode with no login, wallet, stake, or Grid connection.
- Use the built-in decision playground to provide state and questions, then inspect the structured result without sending data off the machine.
- Apple Silicon runs System One with MLX. Windows uses the local CPU runtime.
- Copy the private 127.0.0.1:8765/v1/systemone endpoint for your own local integrations.
- Switch safely between System One, MLX chat models, and llama.cpp models from the same Local Mode catalogue.
- Share Local Insights with four new Singularity Layer visuals for activity, tokens, model use, and generation speed.
- Choose Local Mode for private AI on this computer, or Grid Mode to serve network requests and earn.
- Local Mode needs no login, wallet, or stake. It exposes an OpenAI-compatible endpoint at 127.0.0.1.
- Apple Silicon uses MLX when available. Windows and other supported systems use llama.cpp.
- Fixed the production app policy that blocked the built-in chat from reaching the local endpoint.
- Fixed endpoint controls that could extend outside their dashboard card on smaller windows.
- Use the built-in chat with Markdown, code blocks, tables, links, and a copy button for each message.
- Local Insights records token use, model use, response speed, and activity from the built-in chat.
- Qwen 2.5 Coder on MLX supports OpenCode streaming and tool calls. The app compacts oversized prompts to prevent retry loops.
- Model and context restarts show progress immediately after confirmation.
- Fixed Laya startup on clean machines: the app now provisions its own managed Python runtime when needed, with no Terminal setup required.
- Laya downloads can finish on slower connections without the app giving up, and all English, multilingual, and typed-decision checkpoints are verified before the node joins the grid.
- The Connect flow now immediately shows each startup stage, so longer first-time model setup never looks frozen.
- Simplified onboarding by removing the extra custom-price shortcut. Custom per-model pricing remains available from the Pricing tab after the node is live.
- Added a Settings rollback button for downloading the previous stable app version.
- macOS builds remain Apple signed and notarized. Windows remains a public beta installer.
- Laya now appears as a first-class System One model in the desktop app, so operators can serve it from the Models tab instead of using the CLI.
- The app manages the local Laya sidecar on 127.0.0.1 and starts the normal SGL node background service with the System One runtime attached.
- System One models no longer pretend to be GGUF downloads: there is no model-file download, delete action, or context-window slider for Laya.
- macOS builds remain Apple signed and notarized. Windows remains a public beta installer.
- Windows support is now published through the same release track as macOS, with a one-click installer and signed updater payload.
- Fixed the update dead-end that could leave Windows operators unable to move forward after a node-app update.
- Fixed model-state drift around vision models, so the app no longer warns that you are serving a different model when the running node is healthy.
- The desktop app can keep pulling the latest allowlisted node engine, including the Laya/System One-capable SGL Node v1.9.14, without requiring another app update.
- macOS builds remain Apple signed and notarized. Windows is still a public beta installer and may show Microsoft SmartScreen until Authenticode signing is added.
- The Mac app is now signed and notarized by Apple, so it installs and auto-updates without Gatekeeper warnings.
- Reasoning (thinking) models are better supported on the grid, with the full thinking view.
- Qwen3 32B added to the supported models. It needs a 32GB or larger machine to serve.
- The model list now updates itself from the grid, so new models can appear without an app update.
- Windows 10/11 support is live as a public beta with a one-click installer.
- Windows now has its own updater feed, so Windows releases can move independently without touching the stable Mac release track.
- Auto-update is wired through the Tauri updater and minisign signatures. Install 1.6.0 once; future Windows releases update from inside the app.
- The installer is not code-signed yet, so Windows may show an "Unknown publisher" or SmartScreen prompt. Click More info, then Run anyway.
- GPU acceleration is automatic where available.
- New Reset button on your node: one click cleanly restarts it — handy after an update or if it ever gets stuck — without stopping and reconnecting by hand.
- Node updates now apply on their own. The app restarts your node onto the new version in the background, so you no longer have to restart it yourself for an update to take effect.
- No more surprise password prompt during setup. The node runs models in-process and never needed the separate engine the app was trying to install, so that admin step is gone.
- Signed and notarized by Apple again, so the app opens cleanly with no Gatekeeper warning.
- Your node auto-updates to the latest engine, which makes embedding requests dramatically faster: batches that used to time out now return in seconds.
- Serve embedding models: Nomic, BGE (small/base/large), MiniLM and mxbai now run on the grid via the OpenAI‑compatible /v1/embeddings endpoint — tiny, CPU‑friendly, and perfect for lower‑powered machines.
- Serve reasoning models: DeepSeek R1 (Llama 8B and Qwen 14B) with their thinking shown inline.
- Redesigned Models tab with category tabs — Language, Code, Reasoning and Embeddings — and each model card now shows its provider logo.
- Switch the model you serve in one click: pick a downloaded model and the app cleanly restarts your node on it, with a clear progress state.
- Downloads now show a circular progress ring and can be paused and resumed — big models survive a flaky connection.
- Clearer processing states across Start earning and model switching so actions never look stuck.
- SGL Node CLI v1.7.2 is live: safer Linux GPU serving, better node diagnostics, improved hardware detection, and a real one-slot cap for tight GPUs.
- Run multiple nodes from the same wallet and manage them from the Operator Dashboard on Singularity Cloud Network.
- Recharge credits and use them to deploy private AI Machines or deploy managed grid nodes that can earn without owning hardware.
- AI Machines now hide destroyed/expired instances by default, with a Show archived toggle when you need history.
- Managed grid-node rows now show clearer bootstrap/service state plus node version/build details.
- Vultr AI Machines use the same professional access model as normal Machines: SSH keys when provided, one-time root password fallback when available, and server-only credential handling.
- Headless Linux grid nodes now avoid unnecessary per-heartbeat stake RPC calls; live nodes re-check stake on a cached daily cadence instead.
- If your node is online but its model stops responding (a crash or out-of-memory), the app now shows a clear "Inference engine isn't responding" card with a one-tap Restart.
- Optional Auto-restart: turn it on and the app quietly recovers a stalled engine for you — set it and forget it.
- Prompts you to update to the latest node engine, which stops taking jobs it can't run instead of failing them.
- Dismiss a notification and it stays gone.
- Click anywhere outside the notifications panel to close it.
- The app can now receive alerts meant specifically for your node, not just grid-wide announcements.
- Updates your node to the latest engine that reports its real job status, so the grid never counts a stuck node as free or sends it work it isn't actually running.
- The context-window size you choose is now remembered across restarts and updates, just like your model choice.
- Your Earnings tab now shows on-chain claimable rewards with a one-tap "Claim" button — it opens the dashboard where your wallet signs, so funds go straight to you (the app never touches your wallet).
- A "Verify node" shortcut appears only if your node isn't verified yet.
- Your node now starts with the largest context window your Mac comfortably fits — automatically, no slider needed. More RAM means a higher default; tight Macs stay safe.
- Leaves healthy headroom so the machine is never pushed to the edge.
- You can still set any size yourself — once you move the slider, your choice always wins.
- The model you pick is remembered across restarts and updates — no more reverting to the original one.
- Switching models from the dashboard saves your choice too.
- Context window now scales to your hardware — up to 128K on roomy Macs (was capped at 32K).
- The slider only offers sizes your RAM and the chosen model can actually handle, so no out-of-memory surprises.
- Settings now shows your real installed app version (it was showing a stale number even after updating).
- Your node now updates itself — fully automatic. The app keeps the node engine on the latest version in the background, on launch, with nothing to click and no Terminal.
- When you connect to the grid, the app now quietly updates your node engine to the latest version if it is behind — no prompt, no manual step.
- Added a clear "Save & restart" button for the context-window slider, with an "are you sure?" confirmation — changes apply right away instead of silently later.
- Pick how much conversation your node holds — drag the context-window slider from 4K up to 32K, with a live RAM-impact estimate so you choose what fits your Mac.
- Run longer chats and longer agent jobs without surprises.
- A redesigned dashboard: waveform activity hero, a reputation ring, earnings and throughput charts, and a live jobs feed streaming in real time.
- Model-fit bars show which models run comfortably on your Mac, plus a clear "you keep 80%" earnings readout.
- Switch the model you serve right from the Serving section of the dashboard — no disconnect, no Terminal.
- Launch reward — claim free credits to try the grid, with a scratch-to-reveal card.
- Always On — keep serving with the lid closed (on charger) so your node keeps earning. Toggle in Settings.
- Slimmer, less obtrusive update pill.
- Clearer "Start earning" step when connecting — custom per-model pricing unlocks once your node is live.
- Fixed the first-run tooltip overflowing the window.
- Stop now works for any running node — supervised, background service, or a plain process. No more "stop in terminal".
- A gentle first-run highlight on "Connect to grid" so new operators know where to start.
- Update pill — when a new version is out, a tap-to-update chip appears automatically (and in the notification bell). No more manual checking.
- Delete models you no longer need, right from the app, to free up disk space.
- Fixed Stop — app-started nodes can now be stopped from the app (no more "stop in terminal").
- The sidebar status no longer mislabels your managed node as "terminal".
- One-click install — set up the node engine right from the app, no Terminal and no password.
- Download any model straight from its card; the setup button just becomes "Next".
- Menu-bar icon — see status and pause/resume your node from the macOS menu bar, without opening the app.
- Pricing slider — set your per-model rate visually, with the recommended price marked and a live "cheaper than OpenRouter" comparison.
- A skippable "set your price" step in the connect flow (recommended is auto-selected, so you can just start).
- Account menu in the sidebar — switch wallet or log out.
- Fixed the staking links and the window title appearing white/invisible in light mode.
- One-click Mac node — run confidential AI on the Singularity grid without touching the terminal.
- Smart connect: recommends the best model for your Mac (RAM- and chip-aware) and serves it in one click.
- Runs as a background service — keeps earning after you close the app, and restarts on reboot.
- Custom pricing — set your own per-model rate within the allowed band, right from the app.
- Live earnings, jobs, and reputation. Metadata only: prompts and replies stay end-to-end encrypted.
- Secure auto-updates — every update is signed and hash-verified before it installs.
- Signed & notarized by Apple — opens cleanly, no Gatekeeper warning.
- In-app notifications from the team and a one-tap feedback channel.
The app updates itself automatically: every release is signed and hash-verified before it installs.