Our internet connection has been mucky lately and a few weeks ago it gave up properly. The connection went on Monday afternoon, came back briefly that Wednesday, then dropped again when a power outage kicked in. By Thursday I was relying on data but then the cell towers near me had gone down for a stretch during one of those power outages, and I'd been working off and on for two days without proper connectivity. None of this would normally be a great week, but I got more done than I expected to, and most of that comes down to having a local LLM setup that runs without the internet.

The internet isn't as steady as we treat it

And neither is the power (in my region, at least)

None of this is unique to my region, though my version comes with some extra wrinkles. Power outages come and go here, sometimes they hit every day on a rotating schedule, and sometimes even longer when the schedule slips or when it's an unplanned outage. And our mini UPS keeps the connection going for about three hours after the power drops. Beyond that, it's mobile data, which I'd rather not burn through for what should be quick desktop work.

But the broader thing - the reason any of this is worth writing about - is that an internet connection isn't as reliable as people pretend it is, even without power outages in the mix. For example, friends who travel for work talk about Airbnbs where the router needs a reboot and the host isn't responding. Hotel WiFi often gets throttled in the evenings when everyone's back in their rooms too. Most knowledge workers hit some version of this regularly, and most of the time they just lose the afternoon to it. That is me as well when our internet provider has been as unreliable as it has lately.

My desktop stack does most of the offline work

And it's more than just a local LLM

When our connection cut, I had several PDFs open from a research thread plus a design project I was halfway through. Normally I'd be paging through those in Claude or NotebookLM, but neither was reachable. So I pointed my local LLM at the PDFs and kept going. My runners already have document analysis built-in, and Qwen 3.5 9B handles the synthesis well. And my favorite, Gemma 4 E4B, handles visual work like a champ thanks to its advanced image analysis. The only thing missing was a polish or workspace that a more frontier model would have brought (which I can't run on my limited hardware). But luckily I already had my own local workspace.

Going beyond the LLM, most of my desktop stack already works offline. Obsidian, where all my notes live, keeps everything as plain markdown files locally so it doesn't even notice when there's no connection. My research PDFs also live locally which I can easily attach to any local runner or also open in Obsidian. Besides, I already have my local LLM hooked up to Obsidian through the Copilot plugin, so I can just keep working on notes and call the AI as needed without leaving the app.

Lately, I also added AnythingLLM to my stack which actually gives me persistent memory, which is better than the setup I had before (uploading "journals" of context in the form of plain text documents to my local LLM, and manually keeping them updated). I also recently set up Fabric with my local LLM and Obsidian to speed up longer tasks like summaries for projects with specific briefs - this is the latest addition and I'm still playing around with it, but it works fully offline.

Mobile is the smaller half of the offline workflow

When there's no internet and cell towers are down too, I still had something to fall back on

When we lose our connection, I can still continue working on my desktop without issue - the two days were an unreasonable stretch and I'd much rather be able to use my online toolkit, but my local AI and notes stack held up fine. The problem comes when a power outage hits at the same time, then I'm forced to rely on my phone and Chromebook (which can't run much beyond some browsing and productivity tasks). It gets even worse when the outage stretches long enough for it to affect the cell towers, because then even data isn't an option.

And this is, once again, where a local LLM comes in. Over the past six months or so they've become increasingly accessible and a big part of that is mobile - there are properly optimized models for phones now and runners to match. My main tool is PocketPal with Gemma 4 E2B loaded in it. It's a much more limited setup than the desktop one for sure, but it's more than usable. This model is small by design, meant to run on phones, and it shows - response quality drops compared to Qwen 3.5 on the PC and longer multi-file work isn't really a phone activity in the first place.

But what it does cover is the chunk of small queries I'd otherwise hand to ChatGPT or Claude. I'm talking quick lookups and synthesis of information, anything where you'd quickly pull your phone out to ask a chatbot or search engine. This way, I was able to keep doing research for work and designs and take notes as I was going. When the power (along with the internet) eventually came back on, nothing truly stalled since that Monday, and I could just pick up where the offline work had left off, rather than rewinding to where the connection dropped.

Local LLMs aren't really necessary, but they were in this case

Most of us can skip local LLMs and be completely fine without them. They're not actually necessary, but we use them anyway because of the local and privacy benefits, advanced customization, even the novelty of them, and sometimes offline usability. In my case, they became invaluable when I went for a long stretch of time without connection to the online world.