Free, private AI for the Windows PC you already have Your models, now agents.
- Big models, small memory. A 20B-parameter model runs in just 1 GB of memory.
- Every model, an agent. The models on your PC work your files, websites, apps and tools for you.
What Laghu does
From a model that talks to one that works.
An AI model on its own can only talk. Laghu takes the models on your PC and gives them the means to act, on your own computer, while you stay in charge of every step that matters.
-
1The models you have
Any model, from anywhere
From Laghu’s library, the ones you already have in Ollama, or any .gguf file. No graphics card needed.
- Qwen2.5
- Qwen3
- Llama 3
- Mistral
- DeepSeek R1
- gpt-oss
- Qwen2.5-VL
-
2Laghu
Gives them hands
- Files and commands in a folder you choose
- Websites and Windows apps, with Computer
- Your own tools, through MCP
- PDFs, Word, PowerPoint and Excel, read and written
-
3On your PC
Work done, privately
It reads, writes, clicks and reports back, and asks before anything you can’t undo. The model, your files and your chats stay on your computer.
Agent mode works with models trained to use tools, such as Qwen2.5, Qwen3 and Llama 3.1. Computer works with any model, and best with one that sees. The agent · Computer use · MCP tools
01How it feels
Nothing to set up.
If you can install an app, you can use Laghu.
-
Install it
One download, one installer. No administrator rights, nothing else to set up first.
-
Choose a model
Pick one from the built-in library, or use one you already have. You see its size before anything downloads.
-
Ask
Type the way you’d talk. The answer is written right there, on your own computer.
02Memory and speed
More memory, faster answers.
You decide how much of your computer’s memory Laghu may use. Give it a little and your other apps keep their room. Give it more and replies arrive sooner. Not sure? Leave it on Automatic, and Laghu picks a good setting each time.
How much memory Laghu may use for the model. Auto chooses the fastest setting your free memory allows.
Drag the sliderto see how speed follows memory.
Measured on an everyday laptop
AMD Ryzen 5 5500U with 7.4 GB of memory, of which about 1.4 GB was free.
| Model | Size | Fits entirely in | Speed we measured |
|---|---|---|---|
| gpt-oss 20B | 12.1 GB | 12.2 GB | 0.9 tokens/s with 1.0 GB |
| Qwen2.5 1.5B | 0.9 GB | 1.0 GB | 7.0 tokens/s with 0.9 GB |
| Qwen2.5 3B | 1.8 GB | 1.9 GB | 2.4 tokens/s with 1.0 GB |
| Qwen2.5 7B | 4.4 GB | 4.6 GB | 0.55 tokens/s with 0.8 GB |
| Qwen2.5-Coder 7B Laghu Boost | 4.4 GB | 4.6 GB | 1.44 tokens/s with 1.5 GB |
| Qwen3.5 9B | 5.1 GB | 5.4 GB | 0.46 tokens/s with 0.7 GB |
- Fits entirely in
- The memory a model needs to fit completely, with no risk of running out. Laghu runs it with less; more memory simply makes it faster.
- Tokens per second
- How quickly the answer is written. A token is a word or part of one. Your speed depends on your computer.
03Privacy
What you type stays on your computer.
No account to make. No subscription to pay. Nothing uploaded, to us or to anyone else.
Once a model is on your PC, Laghu works with the internet off. It goes online only when you ask it to, to download a model you chose or to search the web when you allow it, and to check for updates once a day without sending anything about you. You can turn that off, too.
04What you can do
Everyday help, kept private.
Anything you’d ask an assistant, including the things you’d rather not send to one.
Help me say no to my cousin’s wedding, kindly. We can’t travel that week.
Here’s a warm way to put it:
Dear Priya, thank you so much for thinking of us. We’re so sorry we can’t be there on the 14th, but we’ll be raising a glass to you both from home. With all our love, Sam and Leela
Turn these notes into a to-do list, with who does what.
- Send the revised quote to Ana, you, by Friday
- Book the room for the March workshop, Dev
- Share the updated budget with the team, Ana
My electricity bill has a “standing charge”. What is that?
It’s a fixed daily fee for being connected to the supply. You pay it even on days you use no electricity at all, and it helps cover the cost of the network and meter readings.
Why does =VLOOKUP(A2, Prices, 2) give me the wrong price?
Add FALSE as a fourth part, so it looks for an exact match:
=VLOOKUP(A2, Prices, 2, FALSE)
Without it, Excel settles for the closest match it can find, which is why the price comes from the wrong row.
What plant is this? Is it safe for cats?
It looks like a peace lily. It isn’t safe for cats: chewing the leaves can irritate their mouth and stomach. Keep it out of reach, or swap it for a cat-safe plant such as a spider plant.
Rename the photos in Pictures\Lisbon by the date they were taken.
Laghu would like to rename 48 files in Pictures\Lisbon.
05Agent mode Beta
An AI agent that does the work, on your own PC.
Agent mode is in beta. Switch it on and Laghu doesn’t just tell you how. It reads your files, carries out the steps, writes the document and tells you what it did. The AI model doing the work runs on your computer, so no cloud service sees your files.
-
Works in the folder you choose
Give it one folder. Its files and commands stay inside it. Seeing your screen or using other apps is a separate switch, off until you turn it on.
-
Asks before it acts
Reading in that folder is the only thing it does without asking. Changing a file, running a command or going online asks you first, unless you’ve said it always may.
-
Checks its own work
It doesn’t take its own word for “done”. After changing files it runs them, and a command that fails sends it back to fix the cause.
-
Makes real documents
It reads PDFs, Word, PowerPoint and Excel files, and writes PDFs, Word documents and presentations.
-
Shows its sources
When you let it search the web, it links only to the pages it actually opened.
-
Connects to your tools
Add any Model Context Protocol (MCP) server, on your PC or on the web, from a ready-made list or by pasting the config you already use. Every action goes into an activity log you can read. MCP guide.
Total the March invoices in this folder and write a one-page summary I can send to Meera.
- Read 12 invoices in Invoices\March
- Added up the totals with a short script
- Waiting for you
Laghu would like to create March summary.docx in Invoices\March.
06Computer use New
Hand it a task. It uses your computer to do it.
Switch on Computer and Laghu works websites and Windows apps for you. It opens the page, fills in the form, clicks through the steps and tells you what it found, one checked step at a time. The model that does it runs on your own PC, so what it sees on your screen stays there.
-
Looks before every step
Each step starts from a fresh look at the screen: everything on it that can be clicked or filled in, and, for a model that sees, a picture with each one numbered. Then it acts, and looks again to check.
-
Websites and Windows apps
Forms, searches and pages of results in its own browser window, and Windows apps through the same accessibility features screen readers use, so it presses the real button.
-
Never types your secrets
Passwords, card numbers and one-time codes are left to you. It pauses, you type them in, and it carries on.
-
Asks before anything that matters
Paying, sending, submitting, deleting, booking or agreeing to terms waits for your yes, every time. A no stays a no.
-
You can always see it, and stop it
A soft blue light runs round your screen while it works, and turns gold when it needs you. Press Ctrl+Alt+End to stop it from any window. When you use the mouse, it waits.
-
Patient by design
Speed matters less here than getting each step right, so it takes the time a careful person would. Give it the errand and get on with your day.
On the clinic’s website, book the first free Saturday morning slot for me. Stop before you confirm.
- Opened the clinic’s booking page
- Chose Saturday, 9:40, the first free slot
- Filled in your name and phone number
- Waiting for you
Laghu needs your OK: click Confirm booking?
07Requirements
What you need.
- System
- Windows 10 or 11, 64‑bit
- Computer
- An everyday laptop or desktop. More memory makes answers quicker.
- Space
- About 175 MB for Laghu, plus the models you choose
- Graphics card
- Not needed. An NVIDIA card is used automatically if you have one.
- Models
- The built-in library, the ones you already have in Ollama, or your own .gguf files
- Internet
- Only to download Laghu and the models you choose, and for web tasks you give it
- Permissions
- No administrator rights
08Questions
Good to know.
Is it really private?
Yes. Laghu runs on your computer, and your conversations stay there. There’s no account, and nothing you type is sent to us or to anyone else.
Does Laghu work offline?
Yes. Once a model is on your PC, Laghu works with the internet switched off. It goes online only when you ask it to: to download a model you chose, to search the web when you allow it, or to open the websites a task needs when you switch on Computer. It also checks for updates once a day, and you can turn that off.
Can Laghu use my browser and fill in forms for me?
Yes. Switch on Computer in the message box and describe the task. Laghu opens the websites it needs in its own browser window, fills in forms and clicks through the steps, and it works with Windows apps too. It asks before anything that pays, sends, submits or deletes, and it never types your passwords or card numbers: it pauses and lets you do it. Read the guide.
Does Laghu work with MCP servers?
Yes. Agent mode connects to Model Context Protocol servers on your PC or on the web, and their tools, documents and prompts join the agent. Add one from the ready-made list, by address, or by pasting the config from another app. Every MCP tool asks before it runs. Read the guide.
Can I run a large AI model with only 8 GB of memory?
Yes. Laghu runs models that are bigger than your computer’s free memory. On an everyday laptop with 7.4 GB of memory, we ran gpt‑oss 20B, a 12 GB model, with just 1 GB set aside for it. It answers more slowly than a big machine would, and giving it more memory makes it faster. Read the guide.
Do I need a powerful computer or a graphics card (GPU)?
No. An everyday laptop is enough, and no graphics card is needed. If you have an NVIDIA card, Laghu uses it automatically. The more memory you can give Laghu, the faster it answers.
How do I run an LLM locally on Windows?
Install Laghu, pick a model from its built-in library, and start typing. It takes a few minutes, needs no graphics card and no account, and works offline once the model is downloaded. The step-by-step guide.
How much memory should I give it?
Leave it on Automatic, and Laghu chooses a good amount each time a model starts. It never takes more than your computer has free. If you’d like quicker answers and have memory to spare, slide it up.
Which AI models can I run?
Open models such as Qwen2.5, Qwen3, Llama 3.1 and 3.3, Mistral, DeepSeek R1 Distill and gpt‑oss, from small ones that answer quickly to 70 billion parameters and beyond. Bigger models know more and answer more slowly. Pick one from the built-in library, use the models you already have in Ollama, or open your own .gguf files. What each model is good at.
Can I chat with my PDFs and documents without uploading them?
Yes. Attach a PDF, Word, PowerPoint or Excel file, a scan or a picture to a message and ask about it: Laghu reads it on your PC, including the text in scans and screenshots, and nothing is uploaded. In agent mode (beta) it can also work through a whole folder of documents and write new ones from them. How it works.
Can I use the models I already downloaded with Ollama?
Yes. Laghu finds the models you have downloaded with Ollama and opens any .gguf file, so nothing needs downloading twice.
What can agent mode do, and is it safe?
It reads and writes files, runs commands, makes documents and searches the web to finish a task, using an AI model on your own PC. It works only in a folder you choose, asks before it changes anything, runs anything or goes online, and records every step in an activity log. It is in beta, so check its work. Read the agent guide.
Is Laghu an offline alternative to ChatGPT?
For many everyday tasks, yes: writing, explaining, summarising and working with your files, with nothing sent online. Online assistants run much larger models on data-centre hardware, so they’re faster and know more. Laghu trades some of that for privacy, and for working without the internet. An honest comparison.
Do I need an account or a subscription?
No. Download it, install it, use it. There’s nothing to sign up for.
Does Laghu work on a Mac or on Linux?
Not yet. Laghu is for Windows 10 and 11, 64‑bit.
Is Laghu open source?
Laghu is free to use, but its own code is not open source. It is built with open-source components, each used under its own licence and listed on our legal page, and it runs open models made by others.
Will updates interrupt my work?
No. New versions download quietly and install only while Laghu is idle, never in the middle of an answer. Every update is signed by us and checked before it installs.
Windows says it “protected your PC”. Is that normal?
Yes. Windows shows this for programs from publishers it doesn’t know yet. Choose More info, then Run anyway. To be extra careful, first compare the file’s fingerprint with the one under the download button.
Start with the computer you have.
For Windows 10 and 11, 64‑bit. No account, no administrator rights.
- Version
- —
- Size
- —
- Released
- —
The first release is on its way. The button becomes a download the moment it’s published.
Check the download
Every copy of the file has the same fingerprint. If yours matches this one, it’s exactly the file we published.
Loading…
To see your file’s fingerprint, open PowerShell and run Get-FileHash $HOME\Downloads\Laghu-*-Setup.exe