Jais-3-7B splits Modern Standard Arabic into 1.22 tokens per word. tiktoken o200k_base needs 1.83. That is a gap on one small sample, not a verdict on model quality, speed or cost.
Fewer tokens per word means shorter splitting. By our arithmetic, tiktoken used about 50% more tokens than Jais-3-7B on the Modern Standard sentences (1.83 divided by 1.22 is about 1.5).
The short answer
- Source: a developer blog post by Aerin Kim on Miraflow (20 October 2026).
- Method: three parallel sentences in English, Modern Standard Arabic, Egyptian and Gulf Arabic. The unit is tokens per space-separated word.
- Result: both Jais tokenizers split these Arabic sentences into fewer tokens per word than tiktoken o200k_base.
- Limit: it is a small individual test, so treat it as a hint and not a benchmark.
What were the numbers?
| Tokenizer (tokens per word) | Modern Standard | Egyptian | Gulf |
|---|---|---|---|
| Jais-3-7B | 1.22 | 1.57 | 1.55 |
| Jais-3-0.9B | 1.48 | 1.71 | 1.70 |
| tiktoken o200k_base | 1.83 | 1.95 | 2.00 |
The pattern is clear. Every tokenizer splits dialect text more than Modern Standard text, and tiktoken is highest in all three columns. The gap is widest on Modern Standard and narrowest on Egyptian.
How was it tested?
The author says she ran a script comparing tokenizers on three parallel sentences per variety. The code and sentences are published, so you can inspect the experiment.
The script tokenizes text only. It did not run the models to generate answers.
How should you read the results?
Three sentences are not a representative corpus. Your own documents may behave differently.
The numbers do not prove better Arabic quality, a faster model or a lower bill, because token prices and execution differ by provider. A general tokenizer in this test is not automatically the tokenizer of the latest GPT or Claude models.
Who should care, and who can skip it?
- Care if you pay per token for Arabic text and want to know how much tokenizers differ. We suggest running the published script on a sample of your own text.
- Skip if you only need to pick a chatbot for daily use. Tokenization alone says nothing about answer quality.
Questions and answers
Does fewer tokens mean a cheaper API bill?
Not necessarily. Prices per token and model behavior differ, and this test did not measure bills. Compare the price per token and the token count together.
Is Jais better at Arabic?
This test shows only that its tokenizer cuts these Arabic sentences into fewer pieces. Answer quality needs a separate test.
Is tiktoken the tokenizer used by the latest GPT or Claude models?
Not automatically. The general tokenizer in this test is not necessarily the one those models use.
Terms in plain language
- Tokenizer: the tool that cuts text into small pieces, called tokens, before a model reads it.
- Token: a piece of text that AI services count and charge for.
- Tokens per word: how many pieces the tokenizer needs on average for each space-separated word.
- Parallel sentences: the same sentence written in different languages or varieties.
Sources
Facts read on 21 October 2026.
