Models

Five models,
one job.

From 150M to 367M parameters. Pick what fits your machine.

Dhad-1 · under training

The models are still under training; each shows its latest trained version.

Earlier Tashih models

Tashih v1

367M parameters

The latest and most accurate. Learner Arabic, dialects and rewriting.

15.12

Tashih 2

367M parameters

The previous large model.

18.95

Tashih A

216M parameters

A balance of accuracy and size.

20.08

Tashih B

183M parameters

Lighter, with a small loss in accuracy.

21.26

Tashih C

150M parameters

The smallest. For limited devices.

22.40

The number is the word error rate on 78 test paragraphs; lower is better.

Comparison

Smaller, and more accurate.

Our smallest model beats an open model seven times its size.

Uncorrected text34.1
Gemma‑1B — 1B parameters23.09
Tashih C — 150M22.40
Tashih B — 183M21.26
Tashih A — 216M20.08
Tashih 2 — 367M18.95
Tashih v1 — 367M15.12

Which one should I pick?

I want the highest accuracy

Tashih v1

I want a balance of accuracy and speed

Tashih A

My device is short on memory

Tashih C

Try before you choose.

Open the playground