Run your own AI · Best local AI for low-end hardware
The best local AI for a modest computer
thecontinuitylayer.com/open/compare/best-model-for-low-end-hardwareChecked 6 November 2026. US dollars. No affiliate links.
Short answer: on 8GB, Gemma 4 E4B. On 16GB, Gemma 4 12B or Qwen3.5 9B. Both read images as well as text, both handle tool calls, and both are Apache 2.0.
If you have 8GB
| Model | Download | Good for | License |
|---|---|---|---|
| Gemma 4 E4B | 4.5GB at 4-bit (Google's figure) | The best first model for an 8GB machine | Apache 2.0 |
| Gemma 4 E2B | 2.9GB at 4-bit (Google's figure) | Phones, old laptops, quick private chat | Apache 2.0 |
| Llama 3.1 8B | 4.9GB | The most-downloaded base for community retunes | Llama 3.1 Community License |
| DeepSeek-R1-0528-Qwen3-8B | 5.2GB | Step-by-step reasoning on a small machine (it thinks out loud, so it is slower) | MIT (built on Qwen3 8B) |
| Llama 3.2 3B | 2.0GB | Very old or very small machines | Llama 3.2 Community License |
If you have 16GB
| Model | Download | Good for | License |
|---|---|---|---|
| Gemma 4 12B | 6.7GB at 4-bit (Google's figure) | The sweet spot for a 16GB laptop: documents, images, writing | Apache 2.0 |
| Qwen3.5 9B | 6.6–7.6GB | All-round everyday model with vision and tool use | Apache 2.0 |
| gpt-oss-20b | 14GB | Reasoning and tool use, with adjustable thinking effort | Apache 2.0 |
| Phi-4 | 9.1GB | Logic and short reasoning tasks in English; short context | MIT |
Keep it fast
- Close the browser and heavy apps first. The memory they free is what the model uses.
- Use the 4-bit build (Q4_K_M), the usual default.
- Keep chats short, or start a new one. Long conversations need more memory.
- Reasoning models such as DeepSeek-R1 think out loud before answering; on slow machines that wait is long.
- No graphics card? It still works on the processor; stay at 4B or smaller for a usable speed.
All 27 models are in the model finder; how to install one is in your first local model; what to upgrade is in the hardware guide.