VLS 0.1.1 Preview 3.4.2

VLS 0.1.1 Preview 3.4.2 brings tools to Gemma 4, a simpler Benchmarks page that keeps your results, and better support for integrated graphics.

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

• Gemma 4 uses tools. Its own way of calling tools is now understood, so "make an Excel table" runs Python and gives you the file instead of printing code. This also works for apps connected through the API.
• Benchmark results are kept. Every run is saved and still there after you leave the page or restart VLS.
• "Follow a chain of variables" no longer scores 0 on Gemma 4. It was thinking through its whole answer budget on that one task.
• Models remember their settings. Context, graphics offload and the other load options you chose come back every time you load that model. "Reset to recommended" puts the defaults back.
• Integrated graphics (Radeon 780M and similar):
  • the memory reading shows the real shared graphics memory;
  • VLS plans each load around what the system can actually give;
  • if a load still runs out of memory, VLS tries again with less on the graphics instead of failing.
• The activity badge counts only what needs you: errors and finished work. It shows 9+ at most.
• Apps connected through the API:
  • an answer cut at the token limit reports "length";
  • Gemma 4's thinking markers no longer leak into answers;
  • memory no longer feeds a model its own question back.

New

• Benchmarks, redesigned. One button, "Test this model", gives writing speed, time to first token and a quick memory check in about a minute. The longer tests sit under "More tests". Tick two runs to compare them side by side.
• A new update window. It shows what's new, real download progress and each step. After the update, it tells you what changed.
• Smoother windows. Pop-ups, menus and the search palette grow and shrink smoothly when what's inside them changes.

VLS 0.1.1 Preview 3.4.1

VLS 0.1.1 Preview 3.4.1 makes chat check its own work, remember the whole conversation, and stay up in long chats.

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

• VLS no longer closes in long chats. A long conversation with Polish letters or emoji could stop the whole service once it outgrew the model's window. It no longer does, and if anything ever goes wrong in one request, only that request fails.
• An unloaded model stays unloaded. It no longer comes back by itself on the next background request.

Chat

• Checks its own work before answering:
  • sums and percentages go through the calculator;
  • files come out in the type you asked for;
  • a chart or diagram you asked for is really drawn;
  • comparisons look at more than one source.
• Reads the pages it finds, and any link you paste, instead of answering from search snippets.
• Clean answers. No more stray "To:" lines at the start of an answer.
• Diagrams draw. Diagrams are repaired before drawing. If one still has a mistake, the model fixes it itself. Drawings fit their frame, with no cut-off labels.
• Smoother to use:
  • switching chats while one is answering no longer stops or mixes them up;
  • Stop before the first word leaves a Regenerate button;
  • Esc closes side panels;
  • chat search ignores accents.

Memory

• Every chat remembers from its first message, and the chat header shows how much it holds.
• The whole conversation stays within reach. Ask about something from the very start of a long chat, a file you attached early on, what you talked about first, or for a summary of the whole conversation, and it answers from the whole chat, not just its last few messages.
• "From earlier in this chat" under an answer shows what it remembered, and takes you there.
• New chats get up to 1M tokens of memory by default.

Also new

• Open attachments right in the chat.
  • PDFs page by page, with zoom and text you can select;
  • pictures with zoom;
  • spreadsheets as tables;
  • Word documents as pages;
  • code and text.
• Our own Frozen glass, on Windows and Linux, independent of system transparency settings. The title bar is glass too, and the slider goes further.
• An update dot on the VLS icon in the notification area, with "Update to Preview X…" in its menu.

VLS 0.1.1 Preview 3.4

VLS 0.1.1 Preview 3.4 makes chat act like an agent, fixes voice in Polish and other languages, and brings Gemma 4 and AMD graphics up to speed.

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.

New

• Chat that acts.
  • Ask it to look something up and it searches, even when you ask casually ("find something interesting online"). It can search again to dig deeper, and it cites what it found.
  • It never claims it has no internet access while web search is on.
  • It draws charts and diagrams on its own when an answer has numbers or steps.
  • It uses the calculator for sums and makes files (PDF, spreadsheets) when you ask.
  • Schematics draw exactly what you asked for.
• Profile pictures. Twelve tofu to choose from in Settings › General › Account. Until you pick one, your name picks one for you.
• One-click first model. Pick a Bonsai size and press Get started: the model and the engine it needs download together, with one progress bar.
• Notifications that don't spam.
  • Everything that happens is kept, grouped by section. Related events become one entry: a model's load, ready and unload is one card.
  • Repeated confirmations collapse ("Settings saved ×5").
  • Errors stay until you dismiss them, with Retry where it helps.
  • The activity icon counts what's new.

Fixed

• Gemma 4: its thinking no longer leaks into answers, chat titles, web searches or voice, and it gets the thinking switch.
• AMD and Intel graphics: the Windows installer now includes the Vulkan engine and uses it when there is no NVIDIA card.
• Big vision models on small graphics memory: when the vision part doesn't fit, it loads on the processor instead of failing.
• Voice mode:
  • Polish letters (ą, ę, ł, ś, ż…) come through intact.
  • Speech is recognised in the language you speak, and answers are read by a voice of that language.
  • The transcript no longer covers the buttons.
  • Code isn't read aloud.
  • "Preparing voice" no longer hangs.
• Frozen glass is real Windows 11 acrylic: blurred, with no frame, and the title bar is glass too. It needs Transparency effects turned on in Windows.
• Smaller fixes:
  • Menus and side panels no longer slide under the title bar on small windows.
  • The Tools menu fits small screens.
  • Button labels no longer cut off the tails of g, j, p and y.
  • The collapsed sidebar's profile picture is centred.

VLS 0.1.1 Preview 3.3

VLS 0.1.1 Preview 3.3 makes VLS easier to start with and stronger once you know it: Easy and Advanced modes, any model from an API, computer use that finishes the job, and a chat that keeps its head.

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.

New

• Easy or Advanced. The first start asks how much you want to see. Easy keeps the sidebar to Chat, Models and Contexts and proposes one model sized for your computer, from the new Ternary Bonsai 2 family (4B, 9B or 27B), with one Download button. Advanced shows everything. Switch any time in Settings › General.
• Any model from an API. In Connect › Providers, add OpenRouter (its free models are listed first), OpenAI, Groq, Mistral, Together, Ollama, LM Studio or any server that speaks the OpenAI API, then press Use for chat. The model answers in every chat with your memory, and opencode and Hermes reach it through VLS too. Keys are kept on this computer, sealed with your Windows account.
• Frozen glass. A new look in Settings › Appearance: the window turns to frosted glass and your wallpaper shows through (Windows 11).
• Settings, tidier. Sidebar, Home and toolbar arrangement now live together under Appearance.
• Your account at a glance. Your name and avatar sit above Settings, marked private and on this computer.

Computer use

• Finishes what it starts. Measured on ten everyday tasks with a 4B model: 10 of 10, median 18 seconds (Preview 3.2: 3 of 7).
• Saves, prints and files in one step. Save as fills the Save, Open and Print to PDF windows for the model; printing to PDF is one action; making folders, renaming and copying files happen directly, without clicking through File Explorer, and never overwrite or delete anything.
• Checks before it says "done". What a task asked for is checked on disk, and a click that changed nothing is not reported as success.
• Safer. Renaming a drive, folder or file is only possible when you ask for a rename; a task that only checks something types nothing and changes nothing; opening a program by its path asks first.
• Opens the right app. Polish and English app names work (Notatnik, Kalkulator, Ustawienia), and a page opens in the browser you already have open.
• The border fits the screen on displays with Windows scaling.

Chat

• Schematics that work. Ask for the schematic of an ESP32-S3, an ESP32 or Arduino board, a 555, a regulator or a relay and VLS draws a checked design straight away, then the model explains it.
• No more fresh starts. A long schematic or a runaway list no longer pushes the conversation out of the model's memory, so the next message is answered in context.
• Stops runaway answers that count the same line over and over.
• Better web search. Queries are about the topic, not the whole sentence, and pages that are only a bot check or a cookie wall are left out.
• Charts and Python, faster. Charts appear right in the chat; Python starts while the model is still writing the code.

Fixed

• The app window comes back after an update instead of waiting in the tray (from this version on: the update to 3.3 itself is still done by 3.2).
• opencode picks up the model VLS is serving, whichever config file it uses.

VLS 0.1.1 Preview 3.2

VLS 0.1.1 Preview 3.2 fixes what people hit first in Preview 3.1: files you attach, diagrams and charts, the island, and the app window.

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

• Attached files come first. Attach a PDF or a slide deck and say "check this", and the answer is about the file, not about what the chat discussed before. Follow-up questions keep using it, and attaching a file no longer sets off a web search.
• No more lost conversations after a tool error. When code fails, the model sees what went wrong and carries on, instead of starting over with a greeting.
• Diagrams draw. Mermaid diagrams with arrows like → or brackets in their labels are repaired before drawing. If one still can't be drawn, the message says so plainly and the Code tab shows what was written.
• Charts show every slice. Pie and bar charts read the shapes small models really write, so a pie of three categories is three slices, not one ring labelled "1".
• The island no longer shows a light bar above it when clicked, and nothing can move it.
• The app window can't slide under its title bar anymore.

VLS 0.1.1 Preview 3.1

VLS 0.1.1 Preview 3.1 is a polish release: a proper app window, an island that behaves, chats that remember, and a long list of 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.

New

• A real app window. The Windows title bar is now part of VLS, the way modern desktop apps do it:
  • Back and Forward;
  • a sidebar toggle;
  • File, Edit, View and Help menus (Alt or F10 to reach them from the keyboard);
  • the Windows 11 window controls, with Snap Layouts on the maximise button.
    The window is now called Velocity Local Services.
• The VLS logo takes you Home.
• ASCII wallpaper, sharper and in colour. Pictures keep their own colours, and outlines are drawn as strokes along them. Choose your picture again to get the extra detail.

Better

• Chats remember much more. A long chat keeps what was said earlier in it instead of losing it after a few messages, even on small models with short windows.
• The island:
  • stays put at the top of the screen and can no longer be dragged around;
  • opens cleanly the first time after start;
  • needs a slightly longer hold at the top edge (adjustable in Settings › Island);
  • its shortcut recovers if Windows drops it.
  • Talk no longer hangs on "Starting the microphone".
  • A spoken task keeps listening until it is done.
  • Chats asked from the island no longer overwrite changes made in VLS.
• Computer use (still a very early preview, and now labelled as one everywhere it appears):
  • opens the file you asked about instead of an empty app;
  • stays inside the document it is working on;
  • stops immediately when you press Stop;
  • types whole sums into Calculator;
  • permission prompts name the real field;
  • Take over really hands you control.
• Your name is asked once, not again on every start.
• Cleaner chat composer: the Tools button says how many tools are on, and the list is in its tooltip.
• Home is edited from its right-click menu or Settings › Sidebar. The Edit button is gone.

Known issues

• Computer use is a very early preview. It misclicks, gets lost, and sometimes says it has finished when it has not. Watch it while it works.

VLS 0.1.1 Preview 3

VLS 0.1.1 Preview 3 lets your local model use your computer, by text or by voice, from a new island at the top of the screen. And writing in chat feels faster than ever.

Install

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

New

Computer use

• Your local model can open apps, click, type and fill in forms. Works with Qwen3.5-4B on a 6 GB graphics card.
• Fast: it reads each window's buttons and fields directly instead of guessing from pixels; a screenshot is the fallback for apps that don't describe themselves.
• Four levels in Settings › Computer use: Watch only, Ask every action, Ask before risky actions (default), or On its own. Computer use is off until you turn it on.
• You stay in control:
  • Esc, Stop, or pushing the mouse into the top-left corner halts it at once;
  • Pause, Take over and Continue;
  • "Allow for this task" on approvals.
• It never types into password fields, always asks before closing unsaved work, and never acts on a window you didn't name.
• Its own calm cursor shows where it's going and what it's doing; your mouse stays yours.
• Pick which screen it works on. Everything it does goes into an activity log.
• Works with the sandbox: take something from the screen, process it in Python, put the result back.

The island

• A small bar at the top of the screen, hidden until you need it. It opens when you:
  • hold the pointer at the top edge;
  • press the shortcut (shown in Settings › Island);
  • drag a file to the top of the screen.
• Sections:
  • Chat: chat without opening VLS, ask about your screen;
  • History;
  • Talk;
  • Contexts;
  • Activity.
• Talk drives your computer by voice: say a task, interrupt any time ("stop", "pauza", "not that one"), answer permission questions with "yes" or "no".
• While VLS works (answers, downloads, model loading, computer use), it shows a live bar with the current step.
• Hides in full-screen apps and games. Fully configurable in Settings › Island.

Writing in chat

• Instant typing even in long chats; a draft kept for every chat.
• Long pasted text becomes a chip with its size in tokens; paste or drop images and files.
• @ to attach a file or pick a context, / commands with keyboard navigation, a token counter as the window fills up.
• Standard shortcuts: Enter to send, Shift+Enter for a new line, Esc to stop generating, ↑ to edit your last message, Ctrl+Shift+C to copy the last answer.

Settings

• Voice & audio: choose your microphone and speaker, test both, set input volume, noise suppression and echo cancellation.

Starting VLS

• Opening VLS from the Start Menu, search or a shortcut always shows its window. Start-with-Windows stays in the background. Starting again brings the running window to the front. No console windows anywhere.

Known

• Computer use and Talk are new: tested with Qwen3.5-4B on Notepad and Calculator tasks. Start with "Ask before risky actions", and check what matters.
• Linux builds compile and package but have had less hands-on testing than Windows; computer use and the island are Windows only for now.

VLS 0.1.1 Preview 2.2

VLS 0.1.1 Preview 2.2 is a small update that fixes two things people noticed right after Preview 2.

Fixed

• No console windows: VLS starts from the Start Menu, the desktop, autostart and the installer without a terminal window. A new windowless launcher (vlsw.exe) starts it; vls in a terminal works as before. Existing start-with-Windows entries switch over by themselves.
• Moving chats from the browser: the app asks once. If you chose "Bring them over" or "Start fresh", it doesn't ask again, even when the window was closed abruptly. Chats you brought over come back by themselves if the window lost them. An empty browser (a new profile or a private window) can no longer replace the copy that holds your chats.
• Clipped text: text shortened with "…" no longer cuts off the tails of g, j, p, q and y, across buttons, lists, cards, the sidebar, chat and settings.

Install with VLSSetup.exe on Windows (unsigned preview: More info → Run anyway), or with the one-line install for Windows or Linux from the README.

VLS 0.1.1 Preview 2.1

VLS 0.1.1 Preview 2.1 fixes what people hit first in Preview 2.

Fixed

• Schematics:
  • A microcontroller with LEDs, buttons and an RGB LED (ESP32, ESP32-S3, Arduino) now comes from a checked design: every LED with its resistor, every button with its pull-up, free pins only, decoupling included.
  • Antennas are drawn as antennas, not as schematics, and a dipole's length is right when the frequency is given as just "433".
  • Large circuits stay readable instead of shrinking to fit; drag to see the rest.
• Thinking: shown live and open after a web search, instead of hiding as a closed "Thought process" while the answer was still on its way.
• Diagrams: Mermaid diagrams use the app's colours instead of near-black slices. Charts must use the numbers from the sources and the text.
• Loops: an answer that starts repeating itself is stopped where the loop began.
• Thinking box: long thoughts scroll and follow the latest line instead of being cut off; short ones show whole.
• Layout: tabs above diagrams and the password row in setup no longer crowd their edges.
• Updates: preview numbers like 2.1 are recognised, so Preview 2 offers this update.

Install with VLSSetup.exe on Windows (unsigned preview: More info → Run anyway), or with the one-line install for Windows or Linux from the README.

VLS 0.1.1 Preview 2

VLS 0.1.1 Preview 2 is the biggest update so far: VLS now opens as its own app, the chat can search, calculate, run code and draw, and every screen was rebuilt.

Install

• Windows: download VLSSetup.exe and run it. It's an unsigned preview, so Windows may show "Windows protected your PC": choose More info → Run anyway. Or use the one-line install from the README.
• Linux: VLS-linux-x64.tar.gz (or the Vulkan build), or the install.sh one-liner.

What's new

An app of its own

• VLS opens in a native window: dark title bar, remembers its size, hides to the tray. Prefer the browser? Choose it in the first-run setup or in Settings → General.
• A new first-run setup: your name and how VLS opens. Chats from the browser come with you into the app.
• A new start-up animation, and an About Velocity page.

Chat that does the work

• Searches the web first for facts that change: parts, prices, news, events. You watch the sources load, and answers cite them [1], [2].
• A precise calculator with units, plus radio and electronics helpers: wavelength, dipole length, path loss, LED resistor, dividers, dish pointing.
• Pinout, wiring, antenna and schematic drawings, made from what the model says connects to what, and checked for mistakes.
• A Python sandbox that makes PDFs, charts and spreadsheets on your computer, with no internet access.
• Save any chart, diagram, table or answer as PNG, PDF, SVG or CSV.
• Details under every answer: tokens, speed, time to first words, memory, sources.
• Find in chat (Ctrl+F), Esc to stop, ↑ to edit your last message.
• Stops by itself if the model starts repeating itself.

Plugins (MCP)

• Add MCP servers with one click (Files, Web pages, GitHub, SQLite, Memory, Time) or your own.
• Other apps can use VLS as an MCP server.

Look inside the model (Studio → Chat lab)

• Replay an answer token by token: how sure the model was, and what it nearly wrote instead.
• The model's real weights, layer by layer, read straight from its file.

Your VLS, your way

• Customise Home: drag widgets, resize them, add new ones (Live model, Notes, Quick prompts, Fun fact and more).
• Customise the sidebar: hide pages, make groups, fold them.
• A local account with an optional password.
• Backups and exports of chats, contexts and settings, plus restore and automatic backups.
• Uninstall from Settings → About, from the Start Menu or from Apps & features. Your models are kept unless you ask otherwise.

Rebuilt from scratch

• Every screen redesigned: continuous corners, calmer surfaces, clearer wording.
• Smooth transitions between pages and chats, numbers and text that animate, lots of small touches.
• Sizes are always shown in tokens.

Fixes

• Chats no longer pick up memory from other chats.
• Files you attach are read with your message, and you can drop whole folders.
• Saved settings show one clear notice.
• Many smaller fixes.

Known

• Linux builds compile and package but have had less hands-on testing than Windows.
• A 4B model can still get numbers from sources wrong now and then. Check what matters.