You do not need a gaming GPU to start. You need a compact quantized model, enough system memory, a narrow job and realistic expectations about speed.
I Tested Local AI (Ollama) on a 10-Year-Old Laptop
Deviceterra tests Ollama on a 2016 HP ZBook in a live, practical setup rather than relying on a specification sheet. The companion guide helps you interpret that experiment without assuming every old laptop will perform the same way.
Visit the Deviceterra YouTube channel ↗What CPU-only local AI can do well
A modern laptop can run compact quantized models for drafting, summarization, classification, extraction and private question answering. The useful question is not whether it can run the largest model. It is whether a small model can complete one repeated job reliably.
CPU inference is generally slower than supported GPU acceleration, but interactive speed can still be acceptable with small models and short outputs. LM Studio officially supports multiple CPU and operating-system combinations, and llama.cpp is designed for broad local hardware support.
Choose the smallest model that can pass
Start with a reputable instruction model in the 1B to 4B range on 8 GB machines and consider 3B to 8B models when 16 GB or more is available. These are starting ranges, not guarantees. Background applications and context length change memory use.
Use a common 4-bit quantization first. Quantization lowers memory demand by storing weights at lower precision. It makes larger models accessible, but it can reduce quality, so compare the exact file on your task.
A smaller model that finishes your workflow is more valuable than a famous model that overloads the machine.
Keep context under control
Context consumes working memory. Ollama documents that its default context allocation varies with available VRAM, and longer contexts require more memory. A model's published maximum is not the amount your laptop should use by default.
Start with short prompts and only the material required for the answer. For long document collections, use retrieval to select relevant passages rather than inserting everything into every prompt.
A practical first installation
- Close memory-heavy applications and confirm available RAM.
- Install Ollama or LM Studio from its official website.
- Use LocalLens to select a compact model with memory headroom.
- Download one model, not five.
- Test five real prompts and record total response time and answer quality.
- Increase model size or context only when the smaller setup fails a defined requirement.
Tasks to postpone
Large coding agents, long multi-document analysis, web-search agents and multi-user serving demand more context, memory and throughput. Official Ollama integration guidance often recommends 32K or 64K context for coding and agent tools, which can be impractical on modest CPU-only machines.
Begin with single-user, short-context tasks. Upgrade hardware only after a measured bottleneck proves the business value.
Do not buy a GPU to explore local AI. Prove the workflow with the smallest viable setup first.
Official resources
- LM Studio system requirements and offline-operation documentation.
- Ollama context-length and FAQ documentation.
- Hugging Face quantization overview.
- The LocalLens hardware advisor and verified model sources.
Find a model your computer can run.
The LocalLens advisor applies conservative memory rules and tells you when the evidence is insufficient.
Run the free compatibility check →This guide is educational and reviewed for practical accuracy. Model software, licenses and hardware support change; verify official sources before a production deployment.