Guia de IA Local
Rode IA de ponta no seu próprio hardware.
Privacidade total, zero custo por token. Do notebook zerado a um LLM rodando em cerca de 15 minutos.
Por que rodar IA localmente?
Privacy
Your prompts never leave your machine.
Zero cost
No API charges. No rate limits.
Offline
Works on a plane. Works in a bunker.
Full control
Pick the model, weights, system prompt.
Hardware necessário
| Model size | Minimum RAM | Recommended GPU | Speed |
|---|---|---|---|
| 3B – 8B | 8 GB | None / integrated | 20-50 tok/s CPU |
| 13B – 27B | 16 GB | 8 GB VRAM (RTX 3060) | 15-30 tok/s |
| 70B (quantized) | 32-48 GB | 24 GB VRAM (RTX 4090 / M3 Max) | 10-20 tok/s |
| 405B | 64+ GB unified | M3 Ultra or 2x A100 | 5-10 tok/s |
Apple Silicon punches above its weight thanks to unified memory. A 64 GB M3 Max runs Llama 3.3 70B faster than most $3,000 gaming GPUs.
Quantização e GGUF
A raw 70B model is roughly 140 GB. Quantization shrinks it by storing weights at lower precision (4-bit, 5-bit, 8-bit) with almost no quality loss for most tasks. The .gguf format is the community standard for quantized models.
A good default: Q4_K_M. It cuts size by ~4× while preserving 98%+ of the quality.
Escolha seu runner
Ollama
One-line install. CLI + API. The default for hackers.
ollama.com
LM Studio
Beautiful GUI. Best for beginners.
lmstudio.ai
Open WebUI
Self-hosted ChatGPT clone. Multi-user.
openwebui.com
AnythingLLM
RAG + docs on top of local models.
anythingllm.com
Quickstart de 15 minutos com Ollama
- 1Install Ollama from ollama.com. macOS, Linux and Windows supported.
- 2Open a terminal.
- 3Run: ollama pull llama3.3 — this downloads a 40 GB quantized model.
- 4Run: ollama run llama3.3 — start chatting.
- 5For a ChatGPT-like UI, install Open WebUI and point it at localhost:11434.
Melhores modelos locais agora
DeepSeek V3.2
Frontier-class open weights at ~20× lower cost.
DeepSeek R1
Open-source o1-class reasoning.
Llama 4 405B
Open weights. Runs on-prem with full control.
Llama 4 Scout 17B
10M-token context on consumer GPUs.
Qwen 3 72B
Best open-weights multilingual model.
Llama 3.1 8B (Groq)
700+ tokens/sec. Free on Groq.