Impulse 0.1.1 Preview 4.2

Impulse 0.1.1 Preview 4.2 brings Impulse to your phone: a new Settings › Phone page connects the Impulse app in under a minute, at home or anywhere.

Install

• Windows: download ImpulseSetup.exe (small, always installs the newest version) or the full offline VLSSetup.exe, or update from inside the app.
• Linux: Impulse-x86_64.AppImage, impulse_amd64.deb, the .tar.gz, or the install.sh one-liner.

Settings › Phone

• Scan and go: pick how the phone reaches this computer and scan the code with the Impulse app. The address is shown under the code to copy as well.
• One check first: a phone signs in with your profile's password. Without one, or with sign-in from other devices switched off, the page shows the single button that fixes it and keeps the code hidden until then.
• Local network: for the same Wi-Fi. If Impulse only listens on this computer, one button opens it to your network and restarts it. Every address this computer has gets its own code, with a hint when Windows Firewall is in the way.
• Tailscale: away from home on your own tailnet. Impulse finds Tailscale, says plainly when it is missing or signed out, and can turn on HTTPS with Tailscale Serve for you.
• Cloudflare:
  • A quick public link with no account.
  • Or your own Zero Trust tunnel with a public hostname, with an optional Access service token carried to the app inside the code. A short guide on the page walks through the Cloudflare side.
  • cloudflared is downloaded from Cloudflare's official GitHub release only when you ask, and only if it matches the published checksum.
  • Tunnel tokens and secrets are kept encrypted, and the page shows only their last four characters.
  • The tunnel can start with Impulse and stops cleanly when it quits.
• Signed-in phones: every phone and browser signed in from elsewhere is listed with its name, how it connected, its address and when it was last seen. Sign one out, or all of them.

Security

• Requests that arrive through a Cloudflare tunnel or Tailscale Serve are never treated as coming from this computer, even though they connect from 127.0.0.1. They always need a password or the key.
• A quick link refuses to start while Impulse runs without its key or while a profile that may be opened from other devices has no password.

Also

• Apps can list your contexts right after signing in, before any model has been loaded.
• Charts are labelled in the language you asked in.

Impulse 0.1.1 Preview 4.1

Impulse 0.1.1 Preview 4.1 is a polish release: answers that use the web better, thinking that does what you set, models that load faster, and a panel that feels lighter.

Install

• Windows: download ImpulseSetup.exe (small, always installs the newest version) or the full offline VLSSetup.exe, or update from inside the app.
• Linux: Impulse-x86_64.AppImage, impulse_amd64.deb, the .tar.gz, or the install.sh one-liner.

Answers

• Search reads the pages properly: names and topic sentences from the pages reach the model, not just the snippets. Large models can add what they know, marked as their own, when the pages are thin (#17).
• Max thinking thinks: a short question with thinking set to Max is reasoned through instead of answered at once (#18).
• Bold stays: short answers keep their inline highlights (#18).
• Weather and sunsets are looked up, never guessed, and a city's capital is named correctly.
• "And in euro?" converts the previous figure at the quoted rate.
• Comparisons drop search results about other products.
• Budgets with an income get the rest and the shares of income right.
• Translations come back in the language asked for.
• Small talk stays small: "ok", "interesting" or "no" get a short reply, never a file or a search, and a greeting fits the time of day.
• Diagrams keep Polish words in Latin letters.

Models

• NVIDIA drivers 600 and newer: their new nvidia-smi header was read as "no CUDA", so models ran through Vulkan, slower, and models for the NVIDIA engine said no card was found. Both headers are read now.
• Faster loading: a model whose drafting the engine refuses loads straight without it the next time, about 7 s instead of 40 s for Ternary Bonsai 2 9B.
• Honest on computers without NVIDIA graphics: models only the NVIDIA engine can read are no longer suggested there, and the message says so plainly instead of showing a raw log (#16).

Lighter panel

• Pages, Models and the phone tab bar switch about three to five times sooner on a slower laptop.
• Theme changes are one smooth dissolve.
• Long chats and long chat lists open in smaller steps.
• The panel paints at once when the service is busy starting.

Impulse 0.1.1 Preview 4

VLS is now Impulse, by Velocity, and Velocity Context is now Momentum. Same app, same data, same address: nothing moves and nothing needs setting up again. Preview 4 also makes a small model do everyday work properly: real files, real charts, right numbers, and a panel that feels lighter.

Install

• Windows: download ImpulseSetup.exe (small, always installs the newest version) or the full offline VLSSetup.exe, or update from inside the app. Unsigned preview: More info → Run anyway.
• Linux: Impulse-x86_64.AppImage (any distribution), impulse_amd64.deb (Debian, Ubuntu), the .tar.gz, or the install.sh one-liner.

Impulse by Velocity

• New name, new look: the Impulse icon and lockup, a short start animation when the app opens, and Impulse in every window, menu and notification.
• Nothing to migrate: your chats, models, contexts, settings, API key and port 11500 stay as they were. The vls command still works and impulse is added; VLS_* settings are still read.

Work that actually gets done

• Files made by the app: ask for a PDF, Word, Excel or CSV and Impulse builds it from the data itself: categories as rows, totals worked out, and a 12-month plan laid out month by month. The answer quotes the file's own numbers.
• Charts that draw: bar, line and pie charts are drawn from the numbers in the chat, the web or your file, without detours through the calculator.
• Maths done right: equations are solved by the calculator and shown as typeset steps; sums and averages from an attached CSV are worked out from the right column. The calculator takes a decimal comma, temperatures and clock times, and the time difference between two cities comes from real time zones.
• Questions about your files: attach a CSV, a spreadsheet or a PDF and Impulse reads the table itself: sums, averages, the highest and lowest with the right date or name, and the change between two periods. The answer is checked against those figures.
• Turn the chat into a file: "make that a PDF" after a plan, a list or a table puts exactly that in the file, and invoices add up their lines.
• Comparisons come back as tables, and "fix my grammar" fixes the text instead of answering it.
• Asks for a name with a pun get ideas, not a refusal.
• Searches only when it helps: writing an email, fixing grammar, summarising pasted text, maths and code no longer go to the web. Small models have web search on by default for questions about the world.
• Answers in your language, including short ones like "thanks!".

A chat you can shape

• Personality wheel: spin through personalities in the message box and see how each would answer before you pick one.
• Contexts in every chat: pick a Momentum context from a deck of cards right in the message box.
• Model cards with each model's icon, what it can do and whether it fits your graphics card.
• On a phone: a floating navigation bar.

Lighter and faster

• Long chats open several times faster and keep typing smooth: the newest messages show first and older ones come in as you scroll.
• Pages, menus and themes switch sooner, and the personality, context and model pickers open at once.
• Nothing runs while the window is idle.

Fixes

• Studio on AMD and Intel graphics picks a build that runs there, and Studio models can be deleted (#14).
• "regex_error(error_complexity)" no longer stops answers (#15).
• Deleting a chat also deletes its memory, and memory left behind by chats deleted earlier is cleaned up.
• Linux: the tray icon on KDE, the window on start, and updates that roll back cleanly if something fails; impulse doctor checks the install.
• Dialogs: a scrolled list no longer slides under the title, and buttons that don't fit wrap instead of being cut off.
• Model pages from Hugging Face show the publisher's badges as links, not raw brackets.

VLS 0.1.1 Preview 3.5.5

VLS 0.1.1 Preview 3.5.5 makes a small local model work like a big one: it searches, calculates and makes files on its own, gets the numbers right, and never leaves you with a broken answer. Plus a big round of new features and fixes.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

A chat that does the work

• Tools that just work, even on a 9B model. VLS holds tool calls to the model's own format, so they come out right, and the model is offered only the tools the moment needs.
• VLS does the exact parts itself. Sums, averages, percentages and currency conversions are calculated, never guessed. A conversion states its rate and where it came from.
• Facts are checked against the sources. Numbers, scores and names that the pages don't support are sent back for a fix, and the bot no longer says it did something it didn't.
• "Compute it in Python" really runs Python, and the result carries into the next step: variables and files from step one are there in step two.
• Stays on topic and in your language. A short "search the web" after a question about Transformers searches for Transformers, and the answer comes in the language you wrote in.
• Big models get a looser leash. Models of 24B+ and provider models can use every enabled tool and plan many steps; small models stay guided. Change it in Settings › Chat › Advanced › Model freedom.

Files that look designed

• PDF and Word are laid out for their content: a budget gets summary cards and a share chart, an offer gets a net/VAT/gross box and signature lines, a CV gets a sidebar, a how-to gets numbered steps.
• Excel: frozen header, money/date/percent formats, a totals row, a filter and a chart sheet. CSV opens in Excel with Polish letters intact.
• "Add a column and make it again" works: VLS computes the new column from your data.
• Math (LaTeX) renders in chat, PDF and Word.

See it beside the chat

• Artifacts panel: HTML pages, drawings, charts, diagrams, tables and code open in a resizable panel next to the chat, live while they are written.
• Charts read "1 200 zł", "12%" and "2,5 mln", use your number format and show values on hover or tap. Python charts appear in the answer in your theme's colours.
• Tables align numbers, keep their header in view and copy straight into a spreadsheet.
• Diagrams and drawings a small model got slightly wrong are repaired before you see them.

Everyday comfort

• Shortcuts: Ctrl+/ shows them all; ↑ edits your last message; Esc stops.
• Slash commands: /search, /pdf, /excel, /chart, /new, /model, and the same commands in your own language.
• Drop a file anywhere in the window to attach it.
• Search chats by what was said, with the match highlighted when the chat opens.
• Folders and tags for chats, synced to your account.
• Compare two models side by side on the same prompt, with speed and time for each.
• Your own accent colour and a choice of fonts, across the whole app.
• A hands-on guide for the first run that points at the real buttons.
• Live tokens per second while an answer is written, and four starter prompts in a new chat.
• /clear can be undone, and every message the app shows you is in your language.

Nothing gets stuck

• Text written before a tool runs stays on screen; it no longer disappears.
• A dropped connection, a restart or a model switch mid-answer ends in a clear message with Try again, never an endless spinner.
• Two windows on one chat no longer garble or duplicate answers.
• The queue counts only other apps' requests, so your own answer never shows as "1 in queue".

Linux

• The app window: when it can't open, VLS now says why and gives the exact command to fix it. The installer offers to install what is missing.
• No tray on GNOME? VLS keeps the window in the dock and tells you how to get the tray icon.
• Python in chat starts in more environments and explains plainly when it can't.
• Voice setup works on Linux without the Windows-only installer.

Fixes from GitHub

• The Back button in Benchmarks and Settings works (#9).
• Switch between answer versions ‹ 1/3 › under each reply (#10).
• Regenerate and Resend search the web again when needed (#11).
• A deleted chat no longer comes back (#12).
• VLS respects the "start with Windows" setting (#13).
• Studio on AMD recommends models that run there, and Studio models can be deleted (#14).

VLS 0.1.1 Preview 3.5.4

VLS 0.1.1 Preview 3.5.4 makes the chat dependable, even on a small model: it always answers, it stays on topic, and it draws when a picture helps.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

A chat you can rely on

• There is always an answer. When searches, steps or a tool run out or fail, VLS asks the model for a final answer from what it has. For a chart, it draws the points it found and says what is missing.
• One question at a time. A request that went unanswered no longer leaks into the next one.
• Follow-ups are understood. "a ile za tego rtx używanego?" searches for the RTX 3060 you were talking about, in your language.
• Clean sources: only the pages the answer used, with no duplicates, and the numbers in the text match.
• PDF and Excel files with Polish letters come out right, and a Python error is fixed and retried instead of ending the answer.
• Thinking that goes in circles is stopped, and the answer comes without it.
• Pictures when they help: "how does it work" gets a flow diagram, a comparison gets a table, prices over time get a line chart. No chart where a plain answer belongs.

Personality

• Five voices you can tell apart from the first message: Sarcastic, Friendly, Professional, Nerd and Chill. Facts come plainly first, small talk is natural, and the bot never repeats the same opening.
• It never talks about its own instructions.
• Small models start on Chill, which suits them best. Your own choice always wins.

Look and phone

• Frozen glass, refined: proper spacing from every edge, corners that fit together, no seams, and readable text on bright wallpapers.
• On a phone: the chat list is a full screen with swipe to pin or delete and pull to sync, the message box never cuts off its buttons, and landscape works.
• Attached files show a preview in the message box before you send them.
• The Models page: voices have their own compact section, and cards no longer stretch.
• The opening animation finishes in one movement.

Fixes

• Settings › Accounts now shows "can sign in from other devices" right after you set a password.
• English labels left in the attach flow are translated.

VLS 0.1.1 Preview 3.5.3

VLS 0.1.1 Preview 3.5.3 is the biggest update yet: chat that remembers, tools that look the part, a personality, accounts that follow you to your phone, and a much better Linux app.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

Chat

• It remembers. Half the window used to be held back for the answer, so in a long chat the model saw only the last few messages. Now the conversation comes first, and every file you attached stays with the chat; the model no longer says it doesn't have them.
• The same file twice works.
• Tool use, redesigned. Steps read like a short story above the answer: what it searched, what Python ran, what it calculated, with sources as little icons and the code one click away. Two other styles, Live Activity and Cards, are in Settings › Appearance.
• Tables are tables. A table comes as a table, not a picture made by Python. Charts, Excel files and SVG drawings come out right far more often, and broken diagrams are repaired before they're drawn.
• A personality. Pick who answers in Settings › Chat. The default is sarcastic, with jokes about the problem (never about you), facts first. Or switch it off and get the model exactly as its makers shipped it. A quick switch sits next to the message box.
• Answers in the language you set, every time, including in voice mode and with tools.
• Send again under your own message, and Regenerate spins while the new answer is written.
• File preview, rebuilt: smooth zoom around the pointer, drag to pan, double-click for 100%, a clean top bar, and a click beside the file or Esc closes it.
• Save as PDF works for long answers too: clean A4 pages with the chat title, page numbers, Polish letters, and page breaks that never cut a line, a code block or a table row in half.
• The whole chat as a PDF (chat menu or Ctrl+P), with your messages or just the answers.
• Attached files show a preview: the first page of a PDF, or the first lines of a text or code file.
• A fair queue. When several chats, profiles or apps ask at once, everyone takes turns and the chat shows its place in the queue. Settings › Chat › Answers at once lets 2 or 4 answers run side by side, each with a share of the memory window.
• Qwen 3.x tool calls are read correctly, including calls written while thinking.
• Gemma 4 community builds think, including ones with an older chat template.

Accounts and your phone

• Your chats live in your VLS account, not in one browser. Open VLS from your phone or another computer on your network, or over Tailscale, sign in to your profile, and the same chats are there, kept in sync.
• Several profiles, each with its own chats, settings, picture and wallpaper, managed in the new Settings › Accounts tab. Each profile decides whether it may be signed in to from other devices.
• From another device, files come from that device. Adding a folder to a context, importing a model or saving a file now uses the browser on the device you're holding, and VLS never opens a window on its own computer for a remote visitor. Signing in from another device always needs the profile's password; apps keep using the API key exactly as before.
• Your existing chats are moved into your profile on the first start, with a backup kept.
• VLS on your phone feels like an app: a tab bar, the message box above the keyboard, sheets from the bottom, and Add to Home Screen.

Look

• Frozen glass in 2.5D: layered glass panels, cards that float, a glass lens that slides between choices, and contrast that adapts to your wallpaper.
• Your own photo as the wallpaper, with blur and dim sliders, or one of the new gradients.
• 14 languages, all the way through, including menus, fun facts, plurals and the tray. VLS asks your name once and uses it.
• Fuller icons, and Home editing no longer overlaps.

System

• Keep the computer awake while VLS is working, always, or never (Settings › General).
• Restarts and updates wait for the old VLS and its engine to close, so a model is never left behind in graphics memory.

Linux

• The engine never outlives VLS: it is tied to VLS, stopped with it, and cleaned up after a crash.
• Start at boot, no login needed, so VLS is ready for your phone as soon as the computer is on.
• The VLS icon in the taskbar, and frozen glass now finds your wallpaper on KDE, GNOME, Cinnamon and Xfce.
• Back and forward in the browser-window fallback too, and install.sh tells you how to get the native window if WebKitGTK is missing.

VLS 0.1.1 Preview 3.5.2

VLS 0.1.1 Preview 3.5.2 makes the new languages actually show up.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

Fixes

• Languages work. In 3.5.1 the translations were left out of the app, so VLS stayed in English whatever you picked. Now Polski, Deutsch, Español, Français, Italiano, Português (Brasil), Türkçe, Русский, Українська, 日本語, 한국어, 简体中文 and 繁體中文 all show up, including the sidebar and the model menu.
• Start from cards on Home no longer stretch into tall empty boxes, and show more of each prompt.

Some smaller texts, like the facts on Home, are still in English and follow in the next update.

VLS 0.1.1 Preview 3.5.1

VLS 0.1.1 Preview 3.5.1 speaks your language, comes to Linux as a real desktop app, and resumes downloads that break off.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

14 languages

• VLS in your language. English, Polski, Deutsch, Español, Français, Italiano, Português (Brasil), Türkçe, Русский, Українська, 日本語, 한국어, 简体中文 and 繁體中文. The first welcome page asks which one, with your system language already picked, and you can change it any time in Settings › General.
• The installer speaks it too, in every language Windows Setup has a translation for.

Linux

• A real desktop app. VLS opens in its own window with the same top strip as on Windows: back and forward, the sidebar button, the menus and the window buttons. Closing the window keeps VLS running in the tray.
• VLS in the tray, with the same menu as on Windows.
• Quits cleanly. Logging out, shutting down or kill stops VLS, its engine and voice the same way Quit does.
• Back and forward with the mouse buttons, landing where you were.
• Uninstall: install.sh --uninstall.

Downloads

• Downloads resume. A download that breaks off picks up where it stopped instead of starting over, retries by itself, and is checked when it finishes. Pause, resume or cancel each one, and two run at a time.

Chat

• Many more chat settings. Answer style, length, emoji, tables and diagrams, and your own instructions, in Settings › Answers.
• Longer answers by default. The longest answer is now 8K tokens, and an answer cut off at the limit can be continued with one click.
• Everything the model makes opens full screen. Tables, charts, diagrams, code and pictures each have an Open button, with zoom and pan.
• Long table cells wrap instead of stretching the table off the screen.
• New chat is a plain plus.

Models

• The quantisation on every model card, and Bonsai builds named so you can tell them apart: Ternary, smallest (1.58-bit), Ternary, balanced and Plain 2-bit.
• Real provider logos on the Providers page.

Activity

• Every notification opens to show the whole message.

VLS 0.1.1 Preview 3.5

VLS 0.1.1 Preview 3.5 makes Bonsai fast on Radeon and Intel graphics on Linux, and fixes how model sizes and types are shown.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

Linux

• Bonsai on Radeon and Intel graphics runs on the graphics card. VLS used to pick the processor build of the engine Bonsai needs unless VLS itself was the Vulkan build, so a Radeon RX 9070 XT ran Bonsai 2 27B at about 7 tokens a second. Now any graphics card that Vulkan can drive gets the Vulkan build.
• Bonsai no longer crashes while loading on AMD (signal 11). The engine Bonsai runs on is updated to the release that fixes it. Older engine builds are fetched again automatically.
• A proper app window when the built-in one isn't available. VLS opens in its own window through Chrome, Chromium, Brave, Edge or Vivaldi, with no tabs or address bar, instead of a browser tab.
• VLS keeps a log on Linux at ~/.local/state/vls/vls.log.

Models

• The right quantisation everywhere. A file published as IQ4_XS no longer shows as IQ3_S, and Q4_K_M no longer shows as Q4_K.
• Sizes in GB like everywhere else. Sizes were in GiB but labelled GB, so a 13.4 GB file showed as 12.5 GB.
• Gemma 4 thinks. Community builds of Gemma 4 now get their thinking switched on, instead of answering straight away with an empty thought.

VLS 0.1.1 Preview 3.4.5

VLS 0.1.1 Preview 3.4.5 keeps your settings safe through updates, brings a wider Home, much faster voice mode, and a new way to choose how hard the model thinks.

Install

• Windows: download VLSSetup.exe and run it (unsigned preview: More info → Run anyway), or update from inside VLS.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

Fixed first

• Updates keep everything. An update could leave the VLS window with its settings forgotten: back to Easy, and saved benchmark results gone. Now:
  • VLS closes its window cleanly before an update or quit, and the installer asks VLS to quit instead of stopping it by force;
  • your mode, theme and sidebar are also kept by VLS itself, so they come back even if the window's storage is lost;
  • the welcome no longer reappears for someone who already finished it;
  • benchmark results are written safely, with a backup copy.

Voice mode

• Answers start about twice as fast. The first spoken word comes 2–2.5 s after you stop talking, down from 5–8 s, with almost no pause between sentences.
• See it getting ready. The top bar shows voice loading with its steps and real download progress, with Cancel, and Try again if something fails.
• Looks right on Frozen glass, on its own solid surface.
• Never stuck:
  • a failed web search, a model error or a voice engine restart in the middle of a call no longer leaves it hanging;
  • talking while it searches starts your new question instead of running two answers;
  • interrupting a long answer no longer delays the next one.

New Home

• Wider, on a grid of three columns on wide screens, two or one on smaller ones.
• Widgets you arrange: the model and how fast it writes, Continue, Memory, Start from, Getting started, and more under "Add widgets". Drag them to reorder, and give them a small, wide, tall or full-row size.

Chat

• Choose how hard the model thinks with a slider, from Instant to Max.
• Composer buttons never get cut off in narrow windows; labels turn into icons as space runs out.
• Files open full screen: PDFs at 100%, and tables large and centred on a solid background.