Useful things a local AI model can do on your laptop without internet

When the wifi is gone, a small AI model that is already saved on your laptop can still rewrite your draft, outline a messy file, pull names and dates out of your notes, and quiz you on what you wrote. What it cannot do is look anything up. It has no live web facts, no current prices, and no citations you can trust tomorrow. So download the model while you are online, ideally the night before, because turning on airplane mode does not download anything.

Imagine you are on a six-hour flight with no working wifi. On your laptop you have a model of the seven-billion-setting size class (called a 7B, a size that fits on an ordinary laptop) already downloaded, and you have a 2,400-word customer note that still opens with the stiff phrase “As an AI language model.” The customer, Northline, is waiting on a delayed product number, and you need a decent first pass while you sit in your seat, not a portal that will not connect.

This post covers the first useful offline work that setup can do. You get four prompts you can paste straight in, two jobs to skip, and a way to leave a placeholder instead of inventing a delivery date. If all you need is a thank-you note and you have wifi, a normal chat plan is still the shorter path.

Offline is a location you packed

Offline means the model file is already on your disk and your prompt never has to reach a company’s server. You downloaded that file over wifi when you set up the stack. Before you leave, run ollama list, or look at the downloaded list in the LM Studio app (LM stands for language model; it is a free program for running AI models), or point llama.cpp at a compact model file that laptop programs can load (GGUF is the name of that file format), in a folder you can see. The line you want is a local name, not a cloud name. Then flip airplane mode on. A 7B model is the laptop-sized animal described in the earlier post on size and memory, and it will happily remix the words you paste in. It will not open Northline’s portal, and it will not fetch a supplier’s new arrival estimate.

If you open Groq in a browser (the hosting company, spelled with a q, not the xAI chatbot Grok), you are using someone else’s server. The same is true for Together, Fireworks, ChatGPT, Claude, Gemini, and the Grok product. It is also true for an Ollama cloud name if you left that selected. The flight does not care what your model file is called, and location is what counts. This post assumes the 7B is sitting in your computer’s memory (RAM, the short-term workspace a computer uses while running a program), not in a browser tab.

The model’s knowledge also stops at the moment it was trained, and the date is written on the model card (the description page on Hugging Face) you saw the night you downloaded it. So treat the 7B as a rewriter and a practice partner for notes you already have, not as a source of news. Read about size and license while you are on wifi. On the plane, read your own file.

Rule of thumb: If the sentence has to survive contact with a browser, wait for the browser.

Four jobs a 7B can finish in airplane mode

These four jobs have one thing in common: the facts you need are already in a file you brought with you. The quality will be lower than a $20 closed chat, but that is a fair trade when the alternative is staring at the first line for six hours.

Rewrite your own draft. Ask the model to keep every number, name, and date, and to change only the voice. Say who the readers are in one line, for example “Northline operations, who already know the product number.” Ban “As an AI” and brochure adjectives too. If $18,400 of delayed cases becomes any other number, reject that pass. Do not ask for a better delivery estimate either, because the model cannot see your supplier.

Outline a messy file. Ask for section headings and one sentence under each, drawn only from what you pasted. You still choose the final heading names. If the model invents a section called “Market context,” delete it. Gaps in the file are allowed to show up as questions, and new research is not.

Extract fields from your own file. Ask for the owner, date, product number, amount, and next action in a table, and tell the model to write “not in file” instead of guessing. The names, the 12 September 2026 date, and the 400 cases all lived in a text file of standup notes called standups_aug.txt. None of it lived inside the 7B.

Quiz yourself. Ask for eight to twelve questions with the answers hidden, then have the model score you against the same notes. Wrong answers send you back to the file. This is especially useful the day before a vendor call.

Two jobs to skip, and two cousins

Skip live web facts, such as “What is Northline’s current service agreement?” or “Did the supplier post a new estimate this morning?” A 7B will answer in a calm, confident paragraph, but it cannot see the warehouse. You would end up sending a wrong Friday date, in writing, to a buyer who already has the real Friday open in another tab.

Skip citations, current prices, and anything that must still be true tomorrow. A made-up web address looks like a source, and a price from the model’s training year looks like a quote. After you land, those sentences turn into a cleanup job you did not plan for. Plan names and token rates (the per-word charge some vendors bill for AI use) both change, so check the vendor’s page the week you buy, in the same way that prices in this post were last checked in August 2026. Do not let a 7B fill in the cell.

Two cousins belong in the same skip pile. One is legal citations you plan to publish, and the other is button labels in apps like Ollama and LM Studio that change between versions. Frozen model files are not this morning’s news feed. If an answer has to match a page you can open later, park it and leave a stub such as “delivery date: confirm on landing” instead of a fake date.

Good offline jobs vs skip

The four yes rows are what you do on the plane. The two no rows explain why the $14 wifi portal did not matter.

Four paste-ready prompts

Run these against a local 7B-class model you already downloaded, and replace each text block marked FILE with your own text. Do not add a request for live web addresses, prices, or citations. The comment lines at the top are a preflight check you do on wifi the night before, not at 35,000 feet.

# Preflight on wifi ( stack). Then airplane mode.
# ollama list
# Confirm the 7B-class tag is local, not a cloud id.
# Do not pull on the plane. There is no plane wifi that will finish 4.7 GB. # 1) REWRITE. Paste your draft only.
You are editing MY draft. Do not add facts I did not write.
Strip openings like "As an AI" and brochure filler.
Keep every number, name, and date exactly as written.
Audience: [who]. Tone: plain, short sentences.
If you would need the web to write a sentence, write NOT IN FILE instead.
DRAFT:
"""
[paste]
""" # 2) OUTLINE. Same file, headings only.
Using ONLY the text I paste, make H2s and one sentence per section.
Do not add a Market context or Sources section unless those words are in the file.
List gaps as questions, not as new claims.
FILE:
"""
[paste]
""" # 3) EXTRACT FIELDS.
From the text I paste, return a markdown table with columns:
owner | date | sku_or_id | amount | next_action | evidence_quote
Use NOT IN FILE for empty cells. No extra rows from memory.
FILE:
"""
[paste]
""" # 4) QUIZ.
From the text I paste, write 10 questions. Hide answers under each question as "ANSWER:".
Ask me to reply first. Then score me against this same text, not against your training.
FILE:
"""
[paste]
"""

Those four prompts share one paste habit. The quotation marks around your text are your quality control, since a fact that is not between the quotes is a fact the model is not allowed to invent. You still have to read the output, because small models sometimes skip instructions. In our story, the first rewrite tried to add “according to industry reports.” You would delete that clause and add the phrase to the cut list for the second pass.

Worked example: 2,400 words at seat 22A

You start with prompt 1. The 7B takes about 40 seconds on a 400-word chunk, so you split the 2,400 words into six pastes. The first pass kills “As an AI language model, I understand your frustration with the delay.” The new opening names product number 8841, the three-week slip, and your supplier contact, and $18,400 stays $18,400. The whole pass takes about eleven minutes, including the reread.

Prompt 2 produces five headings: what slipped, who is waiting, what we still owe, dates we can stand behind, and what we will confirm on landing. The model offers a sixth, “Comparable retailer delays,” which is not in the file, so you drop it. Prompt 3 runs on standups_aug.txt, which holds 47 bullets. It returns the owners, the 12 September 2026 date, 400 cases, and two “not in file” cells for the carrier’s shipping document. Those empty cells are the real win, because they show the model refused to guess.

Prompt 4 fills the last hour, with ten questions on the eight product numbers. You miss two dates that the file already had. You still have no new estimate, but you do have a draft that no longer sounds like a chatbot, an outline, and a table of owners. The $14 portal never connected, and the job that needed the portal was the one you skipped on purpose.

Same stack, cafe with a dead router

Now say you are in a cafe on a Tuesday and the sign by the register says the wifi is down. You have the same 7B, the same four prompts, standups_aug.txt, and a one-page vendor agenda. Extract the owners, outline the afternoon call, and quiz yourself on the product numbers while your coffee cools. Do not ask Groq or Grok for a Llama price, because those are hosts and not files on your laptop. A cafe is only practice, since you can cheat with your phone’s hotspot. If your hand reaches for the hotspot, the skip pile is talking to you, so put the phone face down and finish the rewrite. The local stack is only as offline as your next browser tab.

Offline does not mean uncensored and legal

These are the slips that catch people most often, and each has a simple fix.

• Pulling a 4.7 GB file (GB means gigabytes, a measure of disk space) after the cabin door closes. Download the 7B on wifi, list it, then switch on airplane mode.
• Pasting a topic (“write about warehouse delays”) instead of your own draft. Offline quality lives in the paste.
• Asking for sources, prices, or this morning’s delivery estimate, because the 7B will comply with a paragraph you will have to unwind later.
• Leaving an Ollama cloud name selected because the local 7B felt slow. That quietly changes your location, and the output of ollama list is the map that shows it.
• Pasting a 40-page PDF on top of 2,400 words and wondering why the model forgets the first line. Break the text into chunks.
• Sending the first rewrite without checking. Read the numbers, because small models drop a zero with a straight face.
• Calling Groq when you mean Grok, or treating either one as a file on the laptop.

Summarize, rewrite, extract, then stop

Pick one file you already have, such as a draft, meeting notes, or a crib sheet. While you are on wifi, confirm the local 7B name with ollama list or the LM Studio downloaded list. Then put the machine in airplane mode for 30 minutes and run prompts 1 and 3. Check every number, and write one stub for a fact you refused to invent. The next post covers memory, graphics chips, heat and battery on a laptop, which explains why that 40-second chunk made the fan loud, and what a graphics chip (GPU, the part that does the heavy math for AI) has to do with it. The stack you are practicing on is described in one local setup from start to finish.

Useful offline, nothing heroic

• Download the 7B on wifi first, because airplane mode is not a download.
• Rewrite, outline, extract and quiz, all on your own file.
• Skip live web facts, citations, current prices, and anything that must be true tomorrow.
• Paste the draft, not a topic, break 2,400 words into chunks, and read the numbers.
• Leave a stub instead of a fake delivery date, and check the stub after you land.

Series notes

This is Part 4 of Run open models from scratch (series code OS11). Previous: one local stack end to end. Next: hardware: RAM, GPU, heat, battery. Related: hosted vs download and privacy paste test.

Sources

• Ollama (local runner; confirm the model line is local, not cloud)
• LM Studio (desktop runner with a downloaded list you can inspect)
• llama.cpp (engine under many local apps; point it at a GGUF you already have)
• Hugging Face: Model cards (size, license, cutoff: read on wifi)
• Hugging Face: GGUF (the packed file format many laptop builds load)
• Groq (hosted inference; not xAI Grok; not offline)
• Analytics Made Simple: one local stack and hosted vs local
• Analytics Made Simple: privacy, run it yourself

Merch on Etsy

Open store →

Apache Hadoop T-Shirt View on Etsy → Naive Bayes Coffee Mug View on Etsy → You're On Mute Coffee Mug – Data Science Gift View on Etsy → Git Version Control Shirt View on Etsy →