Running a local model on integrated graphics
Short answer
Integrated graphics share system RAM rather than having dedicated VRAM, so the limit is how much memory the system will hand over — often more than a cheap discrete card, at lower bandwidth.
How integrated graphics change the arithmetic
A discrete card has a fixed pool of fast memory. Integrated graphics borrow from system RAM, so a laptop with 32 GB can allocate more to a model than an entry-level discrete card has in total.
The trade is bandwidth. System RAM is several times slower than a graphics card's memory, and generation is memory-bound, so a model that fits may still be slow.
What to expect
Usually somewhere between CPU-only and a discrete card, closer to CPU-only on older integrated parts. We have not measured integrated graphics, so this page gives you no tokens-per-second figure — scaling someone else's number and calling it a result is the thing this site does not do.
Driver support decides whether Vulkan is available at all. Install the driver from Intel or AMD directly rather than from the laptop maker, whose builds are frequently years old.
Practical advice
Start with Qwen 3.5 2B at 1.2 GB and move up only if it feels quick. The whole catalogue is on the hardware pages with the fit arithmetic worked out per card.
Other ways people ask this
- integrated graphics local llm
- Will a local model work on my work laptop with only integrated graphics?
Related
- Up a levelGuides: local AI, browser privacy and the hardware to run it
- Closely relatedWhat a local model does to your battery
- Closely relatedWhat a local model actually costs to run
- Closely relatedRunning a local LLM without a graphics card
- Closely relatedCan you run a local LLM on an NPU?
- Elsewhere on the siteCan your PC run a local AI model?
- Elsewhere on the siteHardware guide for local AI