Every AI subscription is priced in dollars, and my electricity bill is priced in taka. That math is exactly why I got curious about running local AI models — open-weight models like Llama, DeepSeek, and their cousins, running entirely on my own laptop with no monthly fee, no API meter, and no internet required. After a few weeks of tinkering, here’s my honest field report from Dhaka: what local AI models can genuinely do on modest hardware, and where the dream falls apart.
Why Local AI Models Suddenly Got Interesting
Two things changed. Open-weight models got shockingly good — recent releases keep closing the gap with the big commercial models on coding and reasoning benchmarks (trackers like LLM-Stats chart the arms race). And the tooling became genuinely easy: Ollama turned “compile mysterious C++ for four hours” into “type one command, get a chatbot.”
What You Actually Need
Less than the gatekeepers claim, more than optimists promise:
- 8GB RAM: small models (3-8B parameters, quantized). Fine for drafting, summaries, simple code help. Think capable intern, not senior dev.
- 16GB RAM: the sweet spot for mid-size models — noticeably smarter, still responsive.
- A GPU helps but isn’t mandatory. CPU-only works; you just wait longer. I treat slow generation as a coffee timer.
Install Ollama, run ollama run llama3.2 (or whatever the current small favorite is), and you’re having an offline conversation with your own hardware five minutes later. It still feels faintly illegal.
What I Actually Use Local Models For
- First drafts and rewrites when the internet is down or the topic is private — client work I’d rather not paste into anyone’s cloud.
- Bulk grunt work: summarizing my own notes, generating alt text, cleaning up transcripts. Free beats metered for high-volume, low-stakes tasks.
- Learning: poking at how models behave without watching a usage dashboard tick upward.
Where Local Loses (Honestly)
The frontier cloud models are still simply better — at long documents, subtle reasoning, and code that touches unfamiliar APIs. My daily-driver AI stack is still cloud-first, and I don’t pretend otherwise. Local models also eat disk (each one is several GB), and battery life on a laptop takes a visible hit while generating.
The Budget Math From Bangladesh
Here’s the framing that matters from here: $20/month is a meaningful recurring cost in taka, every month, forever. A local model costs disk space and some electricity. For maybe 60% of my day-to-day AI tasks, the free local model is good enough — and “good enough for free” wins a lot of arguments when currency conversion isn’t in your favor. My setup now: local for volume and privacy, one paid cloud subscription for the work that genuinely needs frontier intelligence. Hybrid, boring, effective.
Quick Answers
Can I run local AI models without a GPU?
Yes — quantized small models run on CPU. Slower, but usable for drafting and summaries.
Are local models private?
That’s their superpower. Nothing leaves your machine — no logs, no cloud, no terms-of-service anxiety.
Will a local model replace my ChatGPT/Claude subscription?
For everything? No. For the repetitive 60%? Quite possibly. Run both for a month and check where your usage actually goes.
