Running Local LLMs on a Budget Laptop
You don't need a $3,000 workstation to experiment with open-source models. If your laptop has an integrated GPU that can borrow system RAM, you can run surprisingly large models. This is how I run 30B+ parameter models on a Core Ultra class laptop with 64GB of RAM.
Why memory bandwidth matters more than VRAM
Dedicated GPUs are fast because they have high-bandwidth video memory. Integrated GPUs share your system RAM, so their throughput is limited by how fast your RAM is. Roughly, a model generates tokens in proportion to memory bandwidth ÷ model size.
- 8B model (≈5GB at Q4): runs comfortably, 15–25 tokens/sec
- 14B model (≈9GB): usable, 10–15 tokens/sec
- 32B model (≈19GB): slow but works, 5–8 tokens/sec
- 70B model (≈40GB): borderline, 2–4 tokens/sec
Get started with Ollama
Ollama is the fastest way to start. It auto-detects Vulkan-capable GPUs and offloads weights to the integrated GPU automatically.
# Check the GPU detection
ollama list
# Run a 14B-class model
ollama run qwen2.5:14bTwo practical tips
- Always use a quantized format (Q4 or Q5) unless you have headroom to spare.
- On hybrid laptops, Windows may route the app to the discrete GPU. Pin Ollama to the integrated GPU in Windows display settings for big models, or leave it on the discrete card when the model fits in VRAM.
Local models aren't a replacement for frontier APIs, but they are private, free, and great for learning. Start with a small quantized model and scale up.