You've heard about ChatGPT and Claude. Now you want to run AI privately on your own computer. This guide walks you through everything from zero — no experience required.

What Is a Local LLM?

A Local LLM (Large Language Model) is an AI that runs entirely on your own hardware — no internet required, no subscription fees, no data sent to servers. It's the same technology as ChatGPT, but running privately on your PC or Mac.

Why run locally?

Privacy: Your conversations never leave your machine. GDPR-compliant by default.
Cost: After hardware, inference is free. No $20/month subscriptions or API bills.
Control: Modify behavior, use any model, no content restrictions.
Speed: No rate limits, no queue, instant responses.
Reliability: Works offline, no service outages, no API downtime.

Local AI in 2026 is genuinely good. Modern 8B models running on a $300 GPU are competitive with GPT-3.5-Turbo, and 70B models on a $1,600 setup match GPT-4 class quality on most tasks.

What Hardware Do You Need?

The short answer: you probably already have enough to get started.

Hardware	What You Can Run	Experience
Any PC with 8GB RAM (no GPU)	TinyLlama, Phi-3.5 Mini	Works, but slow (~2–5 t/s)
PC with 4GB VRAM GPU	Phi-3.5 Mini, Gemma 3 4B	Good speed (~20–30 t/s)
PC with 8GB VRAM (RTX 3070/4060)	Llama 3.1 8B, Gemma 3 12B	Excellent (~50–70 t/s)
PC with 16GB VRAM (RTX 4060 Ti)	Qwen 2.5 14B, Gemma 3 27B	Great (~40–60 t/s)
PC with 24GB VRAM (RTX 4090)	Qwen 3 32B Q4, DeepSeek-R1-Distill-Qwen-32B	Flagship (~30–50 t/s)
MacBook Air M2/M3 16GB	Llama 3.1 8B, Gemma 3 12B	Excellent — Apple Silicon optimized
Mac with 32GB+ unified memory	DeepSeek R1 32B, Gemma 3 27B	Outstanding
Mac Studio M4 Max 128GB	Llama 3.3 70B, DeepSeek R1 671B	Professional grade

Not sure what GPU you have? Use the Hardware Checker →

The Key Concept: VRAM

VRAM (Video RAM) is the memory on your graphics card. It's what limits which AI models you can run at full speed.

Think of it this way: a model's weights are like a book — the entire book must fit in VRAM for the AI to "read" it at full speed. If it doesn't fit, the model spills into regular system RAM, which is 10–50x slower.

8GB VRAM → Run 7–9B models comfortably
12–16GB VRAM → Run 12–27B models
24GB VRAM → Run 32–70B models (quantized)
Apple Silicon → Your entire RAM counts as VRAM (16GB Mac = ~11–12GB effective for AI)

Step 1: Install Ollama (5 minutes)

Ollama is the easiest way to run local LLMs. It's free, open-source, and works on Windows, Mac, and Linux. Think of it as "the app store + runtime" for local AI models.

Windows

If the installer trips Microsoft Defender SmartScreen, or the terminal cannot find the ollama command afterwards, both are normal on Windows and both have narrow fixes — see why SmartScreen blocks unsigned AI installers and "ollama is not recognized".

Visit ollama.com/download and click Download for Windows
Run the .exe installer (no admin rights needed)
Ollama installs silently and appears in your system tray

macOS

macOS may report a downloaded app as damaged when it is merely unsigned; Gatekeeper and local AI tools explains how to verify the build and open it safely.

Visit ollama.com/download and click Download for macOS
Open the downloaded .zip and drag Ollama to your Applications folder
Open Ollama — approve any security prompts in System Settings → Privacy & Security

Linux

bash

curl -fsSL https://ollama.com/install.sh | sh

Verify the installation:

bash

ollama --version
# Should output: ollama version 0.x.x

Step 2: Download and Run Your First Model

Open a Terminal (Mac/Linux) or PowerShell (Windows). Choose the right model for your hardware:

For 8GB VRAM or 16GB Apple Silicon — the best starting model:

bash

ollama run llama3.1

For older/weaker hardware (4GB VRAM or less):

bash

ollama run gemma3:4b

For 16GB+ VRAM — try a larger model:

bash

ollama run gemma3:27b

For advanced reasoning (if you have 20GB+ VRAM):

bash

ollama run deepseek-r1:32b

Ollama will download the model (4–20 GB depending on size) and drop you into a chat prompt (>>>). Type a message and press Enter.

Essential Ollama Commands

bash

ollama list              # See all downloaded models
ollama pull llama3.1     # Download model without starting chat
ollama rm llama3.1       # Delete a model to free disk space
ollama ps                # See which models are currently running
ollama serve             # Start the Ollama API server manually

Step 3: Choose a Chat Interface

The terminal works, but most users prefer a web interface. Here are your options:

Option A: Open WebUI (Recommended for most)

ChatGPT-style interface that connects to Ollama. Requires Docker:

bash

docker run -d -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  --name open-webui \
  ghcr.io/open-webui/open-webui:main

Then open http://localhost:3000. Features: chat history, file uploads, image analysis, voice input.

Option B: LM Studio (Easiest, no Docker)

Download at lmstudio.ai. A standalone app with built-in model browser and chat interface. Best for non-technical users.

Option C: Jan (Offline-first)

Download at jan.ai. Works completely offline, open-source, privacy-focused.

Step 4: Pick the Right Model for Your Task

Different models excel at different tasks:

Task	Best Model (8GB VRAM)	Best Model (24GB VRAM)
General chat	Llama 3.1 8B	Llama 3.3 70B
Coding	Qwen 2.5-Coder 7B	DeepSeek R1 32B
Writing & summarization	Gemma 3 12B	Gemma 3 27B
Reasoning & math	DeepSeek R1 8B	DeepSeek R1 32B
Fastest responses	Gemma 3 4B	Qwen 2.5 14B
Image understanding	Llama 3.2 Vision 11B	Gemma 3 27B Vision

→ Browse all models → to compare specifications, benchmarks, and download sizes.

Understanding Model Names

When you see model names like "Llama 3.1 8B Q4_K_M", here's what it means:

Llama 3.1 — the model family and version
8B — 8 billion parameters (larger = smarter but needs more VRAM)
Q4_K_M — quantization level (compressed version of the model)

Quantization quick reference:

FP16: Full quality, 2x the VRAM — only for large GPUs
Q8_0: Near-perfect quality, 1x VRAM — great if you have enough VRAM
Q4_K_M: Sweet spot — 5% quality loss, 0.5x VRAM — use this by default
Q2_K: Noticeable quality drop — only for very limited hardware

Troubleshooting Common Issues

"Error: model not found"

bash

# Check the exact model name at ollama.com/library
ollama pull llama3.1:8b   # Specify exact variant

Very slow responses (2–5 tokens/second) Your model is likely running on CPU instead of GPU. Check:

bash

ollama ps   # Shows which layers are on GPU vs CPU

If most layers are on CPU, your model doesn't fit in VRAM — try a smaller/more quantized model.

"Out of memory" error Reduce context length or switch to a more quantized model (Q4 instead of Q8).

Ollama not connecting (on Windows) Check the system tray — Ollama should show as running. If not:

bash

# Restart the Ollama service
ollama serve

GPU not detected on Linux

bash

# NVIDIA: check CUDA driver
nvidia-smi

# AMD: check ROCm
rocm-smi

Key Glossary

VRAM: Video RAM — memory on your GPU. More = bigger models, faster inference.
Parameters (7B, 13B, 70B): Model size in billions. More parameters = smarter but needs more VRAM.
Quantization (Q4, Q8): Compression level. Q4 = smaller/faster. Q8 = higher quality.
Tokens: Units of text (roughly 1 token ≈ 0.75 words). Models charge by token or generate tokens per second.
Context window: How much text the model can "remember" at once (4k = 3,000 words; 128k = book-length).
GGUF: The file format used by Ollama and llama.cpp for quantized models.
Inference: The process of running a model to generate a response (vs training, which creates the model).
MoE: Mixture of Experts — a model architecture where only part of the parameters activate per token.

What Can You Actually Do With Local AI?

Here are real tasks people use local LLMs for daily:

Writing & editing:

Draft emails, reports, articles
Proofread and improve writing
Translate documents

Coding:

Explain code and debug errors
Generate functions from descriptions
Review pull requests
Write tests

Research:

Summarize long documents
Extract key information from PDFs
Answer questions about uploaded files (with RAG)

Personal productivity:

Summarize meeting notes
Generate to-do lists from vague ideas
Brainstorm and ideate
Plan projects

Privacy-sensitive tasks (local AI shines here):

Analyze medical documents privately
Review legal contracts without sharing with cloud
Process financial data internally
Handle confidential HR or client information

Frequently Asked Questions

Is local AI as good as ChatGPT? For most everyday tasks, yes. Modern 8B models match or exceed GPT-3.5. 70B models are competitive with GPT-4o on many benchmarks. The gap has narrowed dramatically in 2025–2026.

How much does it cost to run? After the initial hardware purchase, electricity is the only ongoing cost — typically $5–30/month depending on usage and hardware. No subscription fees.

Do I need a gaming PC? Any PC with a discrete GPU (even an older GTX 1080 with 8GB VRAM) can run useful models. Gaming GPUs (RTX series, RX series) are ideal because they have GDDR6 VRAM optimized for this type of compute.

Can I run local AI on a laptop? Yes — Apple Silicon MacBooks (M1/M2/M3/M4) are excellent for local AI thanks to unified memory. Windows laptops with discrete GPUs (RTX 4060 8GB) work well too. Integrated graphics work but are very slow.

Is my data actually private? Yes. When using Ollama with a locally-running model, all processing happens on your own hardware. No data is sent anywhere. Not even telemetry (unless you opt in). This is fundamentally different from all cloud AI services.

How often should I update models? Check for new model releases monthly — the field moves fast. A model released 3 months ago may already have a significantly better successor.

Your Next Steps

Check your hardware → — Find exactly which models your PC can run
Browse all models → — Explore 35+ local LLMs with specs and benchmarks
Calculate savings → — See how much you save vs ChatGPT or Claude API
VRAM Requirements Guide → — Deep-dive into VRAM math
Ollama vs LM Studio vs Jan → — Choose the right tool for your workflow

Next →Installing Ollama