Drag to rotate the memory board. Each block = 1 GB of weights loaded.
Will it fit? RAM calculator
*Speed estimate assumes a modern laptop with ~100 GB/s memory bandwidth. Generation speed ≈ bandwidth ÷ bytes read per token, so smaller quantized models are faster.
The worked example behind the hype
A model with 8 billion parameters stored at full 16-bit precision needs 8B × 2 bytes = 16 GB just for weights. Quantize to 4 bits and it becomes 8B × 0.5 bytes = 4 GB — small enough for an ordinary laptop, with quality loss usually under a few percent on benchmarks.
Tools like Ollama, LM Studio or llama.cpp pull a quantized GGUF file. After that, no network is touched.
Weights map into memory. Apple Silicon's unified memory is why MacBooks punch above their weight here.
Every token is pure matrix math on your CPU/GPU. Zero cost per message, and your code never leaves the machine.
Popular local models (Q4 quantized)
| Model | Params | RAM needed | Good for |
|---|---|---|---|
| Qwen2.5-Coder 7B | 7B | ~5 GB | Coding agents, autocomplete |
| Llama 3.1 8B | 8B | ~6 GB | General chat, summaries |
| Mistral Small 24B | 24B | ~15 GB | Stronger reasoning |
| Llama 3.3 70B | 70B | ~40 GB | Near-frontier quality, needs a beefy Mac/workstation |
Reality check: viral posts about brand-new "free local models that build whole apps" are often exaggerated. Local 7–8B models are genuinely useful for boilerplate, refactors and offline privacy — but they still trail frontier hosted models on hard, multi-step engineering tasks. Verify claims by running the model yourself; that's the whole point of local AI.