The one-weekend plan
From “8 GB, no GPU” to a working setup.
You do not need a new computer to run a language model, write Python or transcribe a meeting. You need the right order of operations. This page is the shortest route through everything the lab has tested.
| Step | Do this | Time | Report |
|---|---|---|---|
| 01 | Learn the memory math. Understand why a 4.7 GB model fits in 8 GB — and what evicts it. | 10 min | Run an LLM on 8 GB |
| 02 | Pick a small model. 1B–4B quantized. Ignore 7B until step 5. | 15 min | Best small LLMs |
| 03 | Install one runtime. Ollama for convenience, llama.cpp when you want control. | 20 min | Ollama vs llama.cpp |
| 04 | Measure it. Record tokens/sec yourself. A number you measured beats a YouTube claim. | 10 min | Benchmark method |
| 05 | Use it for real work. Short drafts, code explanations, summaries. Not long-form generation. | 1 h | Local coding assistant |
| 06 | Offload the heavy jobs. Free cloud GPU for images, long context and training. | 30 min | Free cloud GPU limits |
| 07 | Decide whether to upgrade. Only if a specific task is blocked — not because a benchmark looks nicer. | 15 min | 8 vs 16 GB |
Before you install anything
Three numbers to know
01 / Free RAM
Windows eats 2–3 GB before you start
On the test machine, a clean boot leaves roughly 5.5 GB free. A browser with ten tabs can take another 1.5 GB. Close it before loading a model.
02 / Model size
Quantization is the whole trick
A 7B model at FP16 needs ~14 GB. At Q4_K_M it needs ~4.7 GB. That difference is why this works at all.
03 / Speed
Below ~3 tok/s, chat breaks down
Reading speed is roughly 5–7 tokens per second. If the model generates slower than you read, waiting dominates the task.
Choose your route
Three destinations
LOCAL AI
Chat and write offline
CODING
Ship code, not crashes
HARDWARE
Spend money once
A note on honesty
Your 8 GB is not my 8 GB.
CPU generation, single- vs dual-channel RAM, SSD speed and thermal throttling all change the result. Treat the numbers here as a floor for a 15 W 2018 laptop, not a guarantee. The method page explains exactly how each measurement was taken.
Read the test method