Measured, not estimated
The 8 GB benchmark table
Every model below was loaded and prompted on the same laptop. Numbers are ranges from five runs with the browser closed. If a model needed swap to answer, it is marked accordingly.
| Model | Quant | File | Load | First token | Generation | RAM headroom | Verdict |
|---|---|---|---|---|---|---|---|
| Qwen3 0.6B | Q8_0 | 0.7 GB | 5 s | ~2 s | 25–35 tok/s | ~4.9 GB | Runs |
| Qwen3 1.7B | Q4_K_M | 1.1 GB | 8 s | ~3 s | 12–18 tok/s | ~4.2 GB | Runs |
| Llama 3.2 3B | Q4_K_M | 2.0 GB | 15 s | ~8 s | 6–9 tok/s | ~3.1 GB | Runs |
| Phi-4-mini 3.8B | Q4_K_M | 2.3 GB | 16 s | ~9 s | 5–7 tok/s | ~2.8 GB | Runs |
| Gemma 3 4B | Q4_K_M | 2.5 GB | 18 s | ~11 s | 4–6 tok/s | ~2.4 GB | Runs |
| Mistral 7B | Q4_K_M | 4.4 GB | 38 s | ~90 s | 2–4 tok/s | ~0.9 GB | Needs tweaks |
| Qwen2.5 7B | Q4_K_M | 4.7 GB | 40 s | ~95 s | 2–4 tok/s | ~0.7 GB | Needs tweaks |
| Llama 3.1 8B | Q4_K_M | 4.9 GB | 52 s | ~120 s | 1.5–3 tok/s | ~0.3 GB | Avoid |
| Whisper base | ggml | 145 MB | 3 s | — | 10 min audio ≈ 7 min | ~4.6 GB | Runs |
| Whisper small | ggml | 470 MB | 5 s | — | 10 min audio ≈ 14 min | ~4.3 GB | Needs tweaks |
| SD 1.5 (image) | — | cloud | — | — | 20–40 s / image | local: no | Cloud only |
RAM headroom = free memory while the model is loaded and idle, with no browser open. Below ~0.8 GB, Windows starts paging and tokens/sec collapses.
Do the math for your machine
RAM headroom calculator
Enter your total RAM and what else is running. The calculator suggests the largest quantization that still leaves a usable margin.
Free for a model: about 5.5 GB. Best fit: a 4B model at Q4_K_M (≈2.5 GB) with room to spare. A 7B Q4 (≈4.7 GB) fits but leaves under 1 GB — expect paging if anything else opens.
How these numbers were produced
All tests ran on one machine: a 2018 15 W ultrabook with a Core i5-8250U (4 cores / 8 threads), 8 GB of DDR4 RAM in single-channel mode, integrated UHD 620 graphics and a 256 GB SATA SSD. No external GPU, no eGPU, no overclocking.
Protocol
- Reboot, wait two minutes, close every non-essential app.
- Set Windows power mode to Best performance and connect AC power.
- Load the model with Ollama or llama.cpp, then wait until memory use stabilises.
- Send the same three prompts: a one-line question, a 150-word summary task, and a 40-line code explanation.
- Record time to first token, generation speed for the final response, and free RAM at idle.
- Repeat five times; report the middle range, not the best run.
What “tokens per second” means here
Generation speed only. Prompt processing is reported separately as “first token”, because a long prompt can take far longer than the answer. A model that reads a 2,000-token document in 90 seconds and answers at 4 tok/s is still usable for summarization and unusable for chat.
Thermal reality
This chassis throttles after roughly eight minutes of sustained inference. The first answer of a session is faster than the tenth. Sustained speeds in the table are 10–20% below burst speeds for that reason.
Run the same test yourself
With Ollama installed, this gives you a rough generation speed on your own hardware:
ollama run qwen3:1.7b --verbose "Write exactly 100 words about why SSDs matter."The eval rate line at the end is your tokens per second. Compare it with the table before deciding that a bigger model is a good idea.
Next step
Pick the model, then read the install guide.
The benchmark tells you what fits. The field reports tell you what the experience is actually like — including the parts that are too slow to recommend.