Compare the models

The chat on this site offers four models that run on your graphics chip. They differ mainly in how much you download once, how much memory they need, and how quickly the answers come out.

Try them in the chat Check what this device can run Questions and troubleshooting

Qwen2.5 0.5B Instruct

Default

One-time download380 MB, kept in the browser afterwards.

Memory neededAbout 1 GB free memory.

SpeedThe fastest of the four to download and to answer. Fine on an older tablet.

Best forSmallest and fastest. Good for short answers, rewriting and summaries.

Llama 3.2 1B Instruct

One-time download880 MB, kept in the browser afterwards.

Memory neededAbout 2 GB free memory.

SpeedRoughly half the speed of Qwen 0.5B, and still comfortable on a laptop.

Best forNoticeably better conversation quality for a bigger download.

Qwen2.5 1.5B Instruct

One-time download1.09 GB, kept in the browser afterwards.

Memory neededAbout 3 GB free memory.

SpeedA little slower again than Llama 3.2 1B. Best on a desktop or a recent tablet.

Best forA strong all-rounder. Best on a desktop or a recent tablet.

Phi-4 mini Instruct

One-time download2.3 GB, kept in the browser afterwards.

Memory neededAbout 5 GB free memory.

SpeedThe slowest to start and to answer. Needs a recent desktop or laptop.

Best forThe most capable option. Needs a recent desktop or laptop with graphics acceleration and plenty of memory.

About those speed figures

There is no honest single number here: the same model can be several times faster on a desktop graphics card than on a tablet. The ordering above follows model size, which is what actually decides the difference on any one device. The one figure measured on this site is for processor-only mode, where the smallest model produces roughly two and a half words a second.

How to choose

Older iPad or laptopStart with Qwen2.5 0.5B. It is the quickest to download and the least likely to run out of memory.

An everyday laptop from the last few yearsLlama 3.2 1B is the sweet spot — clearly better conversation for a download you only do once.

Desktop with a good graphics cardPhi-4 mini gives the most capable answers, with Qwen2.5 1.5B as a lighter middle ground.

No graphics acceleration?

If your browser cannot reach the graphics chip, the chat quietly switches to a much smaller model that runs on the processor. Those two are separate from the four above:

SmolLM2 360M Instruct (processor only)270 MB to download. Quickest to download and start. Best choice on an older laptop or iPad.

Qwen2.5 0.5B Instruct (processor only)380 MB to download. Somewhat better answers, a bigger download and a slower start.

The diagnostics page tells you which path your device will take, and what you can change.

More about Llama 3 in the browser Why the Neural Engine isn't used