// measured here, in text — free, open models you run yourself

Then, which one would you rather talk to?

Free models you can run on your own computer, ranked on one question: which of them is worth talking to? Same scenarios, same judge, same instructions for every model — and the whole ladder is below, most human at the top.

How it works· Head to head

How human each one sounds

Every model we have run, on one scale. Higher is more human — the number is a normalised Elo from head-to-head judging, not a percentage.

Loading…

Graded by gpt-5.4, using the same instructions for every model.

//

The full table

Every column we record, for every model. Click a heading to sort.

Speed is in words per second, listed for every machine we measured a model on — a DGX Spark, an RTX 5090 and an RTX 5080. Most people read about 4 words a second, so anything above that is already arriving faster than you can read it. Machines a model is too big to fit on are left out.

▸ Why some rows are the same model twice

Each row is one way of running a model. The same model compressed differently, or run with different software, is a separate row because it behaves differently. Each row links to its full record.

less human more human — color ranks each model within its own column; don't compare colors across different columns.

Loading…