FAQ: Frequently Asked Questions

Answers to the most common questions about local AI models, hardware, coding, and safety.

🇷🇺 Russian version: faq.ru.md

General questions
About hardware
About models
About coding
About safety and legality
Whats next

1. General questions

Is AI dangerous?

Analogy: imagine you saw a car for the first time. It is fast, powerful, can kill — but you are not afraid of every car on the road. The danger is not that the model “wants” to harm you, but that a person might blindly trust its answers. Treat AI as a helper that can make mistakes — and you will be fine.

Is ChatGPT the only AI?

ChatGPT is one specific service from OpenAI. But there are dozens of other models: Llama (Meta), Qwen (Alibaba), Mistral, DeepSeek, Gemma (Google). They can be just as good as ChatGPT — and run entirely on your computer, for free, without internet.

ChatGPT is the tip of the iceberg. Below the surface lies an ocean of open models you can run locally.

Will AI replace my job?

AI will not replace you. But a person who knows how to use AI will replace one who does not.

Just like Excel did not replace accountants but made them more effective — AI is becoming a tool of augmentation, not replacement. Especially in professions that require local knowledge and human contact — AI is still weak here.

Learning AI now is not about fleeing replacement — it is an investment in your competitiveness.

Why run locally if ChatGPT exists?

Argument	ChatGPT	Local model
Privacy	Everything goes to OpenAI servers	Nothing leaves your computer
Price	$20/month or usage limits	Free (just electricity)
Availability	Needs internet	Works offline anywhere
Control	Model can be changed or “fixed”	Your model, your rules
Customization	Cannot fine-tune	Can fine-tune for your tasks

More details: cloud-vs-local.md

2. About hardware

How much does it cost to run local models?

It is free. If you already have a computer, you only pay for electricity (pennies). Models are downloaded for free and run without subscriptions.

The only expense would be if you decide to upgrade your hardware — but that is a one-time investment.

I have 8 GB RAM — can I run anything?

Yes. Run small models (1-3B parameters): Phi-3-mini, Qwen 2.5 1.5B. They work on 8 GB, though speed will be modest (~10-15 tok/s). Enough for simple chat and code autocomplete.

If budget allows — a used MacBook Air M1 16 GB will give a much better experience.

Do I need a GPU?

Not necessarily. Modern tools (Ollama, LM Studio, llama.cpp) work well on CPU. GPU accelerates by 2-5x but is not required.

Exception: for 14B+ models GPU is recommended. For models up to 7B, CPU is sufficient.

Is Mac bad for AI?

Outdated myth. Apple Silicon (M1/M2/M3/M4) is one of the best platforms for local AI:

Unified Memory: all RAM is available to the model
MLX Framework: often faster than CUDA on comparable hardware
Energy efficiency: MacBook runs on battery, PC with GPU consumes 300-450W

Intel Mac is not suitable. But Apple Silicon is an ideal starting point.

My computer is 5 years old — can it run?

Depends on configuration:

5-year-old MacBook Pro (Intel) 16 GB — can run models up to 7B, slowly
5-year-old Windows laptop 8 GB without GPU — only 1-3B models
Tip: use Q4 quantization and small models. Start with Qwen 2.5 1.5B.

3. About models

How many models can I install?

As many as you want. Each model is a file (4-40 GB). Ollama stores them in ~/.ollama/models/.

ollama rm qwen3.5:4b    # remove a model
ollama list              # see installed models

Models give wrong answers (hallucinations). What to do?

This is called hallucination. The model does not know the boundaries of its knowledge — it always outputs the most likely continuation. If a question falls outside its training data, it will still answer, but the answer may be incorrect.

What to do:

Verify important facts
Use RAG (search your documents) — the model answers based on your data
Lower temperature (0 = fewer hallucinations)

Which model is best for Russian?

Qwen 3.5 from Alibaba is the best for Russian among open-source models. It was trained on a significant volume of Russian-language data and understands context, idioms, and cultural references.

Also good: Llama 3.1 (worse, but can be improved with system prompts) and DeepSeek-R1 (excellent reasoning, average Russian).

How to update models?

Models do not update automatically. When a new version is released, simply download it as a new model. The old one can be removed.

Ollama itself is updated via:

brew upgrade ollama  # Mac
# Or reinstall from ollama.com for Windows/Linux

Can I run multiple models at the same time?

Ollama keeps a model in RAM while you work with it. If you start a second one, the first gets unloaded. You can configure keep_alive for both, but this will quickly fill up RAM.

Practical tip: use one model for chat (Qwen 3.5 7B) and one small one for autocomplete (CodeGemma 2B). They fit in 16 GB RAM.

4. About coding

Can I use local models for coding?

Yes. This is one of the main use cases. Install Ollama, download a coding model (Qwen 2.5 Coder 7B), and connect to VS Code via Continue.dev or Aider.

Step-by-step: ../use-cases/coding.md

What is vibecoding?

Vibecoding is an approach where you describe what you want, and AI writes the code. You review, give feedback, AI fixes.

What is Aider?

Aider is a terminal-based coding tool. You give it a task, it reads your codebase and makes changes. Works with local models via Ollama.

More: ../use-cases/coding.md

OpenCode / Cursor do not work with local models — what to do?

For local models use:

VS Code + Continue.dev — plugin, works with Ollama
Aider — terminal coding agent
Cline — VS Code extension with local models
Void — open-source Cursor alternative

More: ../use-cases/coding.md

5. About safety and legality

Is it legal to use open-source models for commercial work?

Yes, in most cases. Major open-source models (Llama 3, Qwen, Mistral, Gemma) have commercial licenses — they can be used in business.

Nuances:

Copyright on generated code is debatable (same as for GitHub Copilot)
Practical tip: check each model’s license on its page
For content — do not pass generated material off as your own without editing

More: ../agents/safety.md

Are local models safe? Can they harm me?

A local model is safer than a cloud one because:

No data leaves your computer
The model has no internet access (unless you give it one)
You control which prompts to send

But: like any tool, a model can be used for harm. Do not ask the model to do illegal things, and do not blindly trust its medical or legal advice.