Guide 🖥️ Run AI on your own PC — the zero-cost local-LLM starter

Every AI chat you've used runs on someone else's computer. They see your prompts, they set the limits, they raise the prices. A local LLM is the same idea running on YOUR machine: free after zero rupees, works offline, and your prompts never leave the room. This is the starter path — no jargon, no GPU cult talk.

✅ The honest gear check

• 8 GB RAM — runs small 3B models fine (okay for notes, summaries, simple chats)
  
• 16 GB RAM — the sweet spot: smooth 7–8B models, genuinely useful
  
• Any gaming GPU (6–8 GB VRAM+) — everything gets 3–5x faster, but it's a bonus, not a requirement
  
• Apple Silicon Mac (M1/M2/M3/M4) — weirdly excellent at this; 16 GB unified memory sings
  
• Disk space: 5–15 GB per model (they're just big files)
  

No "you need an RTX 4090" nonsense. If your PC runs Chrome and a game, it can run a small model.

⚡ Path A: Ollama (recommended)

Two minutes to first chat:

Code:

1. Download from ollama.com (Windows / Mac / Linux installer) 2. Open a terminal and run: ollama run llama3.2 3. It downloads the model (~2–5 GB), then you're chatting.

That's the whole install. Type normally, ask things, it answers. Offline forever after the download.

🖼️ Path B: LM Studio (the pretty one)

• Download from lmstudio.ai — full graphical app, no terminal
  
• Search models inside the app, click download, click chat
  
• Pick this if you hate command lines — same engines underneath
  

Rule of thumb: Ollama if you'll build things later, LM Studio if you just want to play today. Both free.

🧠 Which model first

• llama3.2 (3B) — fast everywhere, decent chat, start here
  
• qwen3 (8B) — the 16 GB people's favorite, strong all-round
  
• mistral (7B) — tidy writing, good with structure
  
• deepseek-r1 (8B) — reasoning-style answers when you need to think through a problem
  

Big number rule: B = billions of parameters. Bigger = smarter but slower and hungrier. When in doubt, run the smaller one quantized — a fast good answer beats a slow perfect one.

🐢 "It's slow!" — the 3 usual fixes

• You picked too big a model for your RAM — drop from 8B to 3B and watch it fly
  
• Close Chrome, Discord, and your game launcher — they eat the RAM the model wanted
  
• Use the quantized (Q4) version — LM Studio lists them; almost same smarts, half the size
  

On CPU-only, expect reading-speed answers, not instant ones. Normal. GPU owners get spoiled.

🔌 Make it useful, not just a toy

• Open WebUI — gives your local model a ChatGPT-style interface, chat history, file upload. One Docker command and it talks to Ollama automatically
  
• Local API — Ollama runs a mini-API on your machine at port 11434, so your scripts (Python, n8n flows) can call YOUR model with zero API bills
  
• Private documents AI — feed it your PDFs and ask questions with citations. Full build here: Talk to your PDFs — private document AI
  

🚫 When local is the WRONG choice

• You need the absolute smartest answer on a hard problem (cloud frontier models still win)
  
• You're on a 4 GB laptop from 2014 (be kind to it)
  
• You need image generation or live web search (possible locally, but that's week-two territory)
  

The honest pitch: local AI won't replace your cloud subscription today. It WILL handle the everyday 80% — summaries, drafts, private notes, experiments — for free, offline, and private. Learn the muscle now; the models only get better from here.

Set it up, then come tell the forum what your PC managed to run — benchmarks welcome below. 👇