WHAT NIGHTJAR CAN DO

Local AI, without the afternoon of setup.

One download. No terminal, no build flags, no account, no API key. Pick a model, then ask about whatever you have open. The model runs on your own machine, so the pages you ask about, the files you drop in and the questions you type are never sent to an AI service. Open the panel with the Ask Nightjar button at the top of the window.

IT LOOKS AFTER ITSELF

What it does for you

It sets itself up

Automatic

Running a local model normally means choosing a build for your chip, picking a quantisation, and guessing how many layers to put on the GPU and how much context your memory will take. Get that last one wrong and nothing errors. It just runs badly, and you cannot tell whether the model is poor or your settings are.

Nightjar sizes the context and the GPU layers to the memory you actually have, and labels every model fast, usable or too large before you download it.

It tells you what your model is good at

Measured

The first time you load a model it takes a short reading test with known answers, from an article, a spreadsheet and a PDF, and reports where it is strong and where it is not. No guessing whether a smaller model is good enough for what you want.

THINGS TO TRY

Ask, drop in, calculate

Ask about the page you're on

Reliable

Switch on the 📄 Page chip above the chat box and Nightjar reads the current page and answers from it, citing the part it used. Works on articles and PDFs. Ask something that is clearly about the page while the chip is off, and it asks whether to use the page rather than answering blind.

Summarise this article in five bullet pointsWhat is the main argument here?What does this page say about refunds?

Read a document you'd rather not upload

Reliable

Drag a PDF, text file, CSV or Excel workbook into the chat and ask about it. It is read on this machine, which is the point: contracts, statements and anything else you would not paste into a website.

Summarise this documentWhat are the key dates in here?

Get a calculated answer from a spreadsheet

Calculated

Ask a CSV or Excel file for a total, an average, a count, the highest or lowest row, a difference or a single value. The model only turns your question into a query; Nightjar works the answer out over every row and shows its working, so a small model's arithmetic is never the answer.

When it cannot calculate something, it says so before the answer and says why, and the model's own attempt is marked as one to check.

What was the total revenue for the North region?Which month had the fewest faults?How many orders were over 3,000?

Ask it anything else

Reliable

It is a general assistant that happens to live in a browser. Questions, explanations, drafting and rewriting all work with no page at all. Code in an answer comes in its own box with Copy, and Format for JavaScript, JSON, CSS, HTML and the like.

Draft a polite reply declining this invitationExplain how DNS works
WHAT TO EXPECT

Where it shines, and where it doesn't

Works well

  • Questions about the page you're on
  • Summarising long articles and PDFs
  • Explaining and rewriting text
  • Reading documents you drop into the chat
  • Totals and counts over a spreadsheet you attach
  • General questions with no page open

Hit and miss

  • Arithmetic over a table on a web page, or inside a PDF
  • Figures inside charts and images
  • Very long tables that load as you scroll
  • Pages behind logins or heavy pop-ups
  • Anything depending on very recent facts

A model that runs on your own hardware is far smaller than one in a data centre. That trade buys privacy and offline use. Where a page cannot be read properly, Nightjar says so rather than inventing an answer. If it tells you something is not on the page, that is a real answer, not a failure.

MODELS

Choosing a model

Start small, move up if you need to

The Models tab lists the catalogue and flags anything too large for your hardware before you download it. Qwen 3.5 2B (1.2 GB) is a good first choice on any machine. You can also import any GGUF file you already have, or use models from Ollama or LM Studio.

Bigger is slower, not always better

Bigger models reason better but respond more slowly. If replies feel sluggish, try a smaller one, or switch between GPU, CPU and hybrid. On Nightjar's own reading test the 1.2 GB Qwen scored full marks where some larger models did not.

PRIVACY

Your pages, files and chats never go to an AI service

Page content, attached files, chat messages and typed text are processed by the model running on your own hardware. Ads and trackers are blocked as you browse, so the pages you visit leak less too.

Downloading a model, the filter lists, a once-a-launch check for a newer version and the pages you choose to visit are the network traffic you would expect. All of it is listed in the privacy policy.

Open the panel with Ask Nightjar · pick a model · ask. More help: help and documentation · install steps