A vocabulary is a budget — Sarvam-1 divided its 68,096 slots like this

Latin 26.4%

Devanagari 17.8%

8.7

8.4

7.8

7.7

7.5

7.0

5.3

2.9

Englishand code

Hindi, MarathiDevanagari

Telugu

Kannada

Bengali

Gujarati

Tamil

Gurmukhi

Malayalam

Oriya

Where all 68,096 slots go

learned tokens — the bar above63,741

reserved — room to grow later4,096

raw byte values — the fallback net256

special — unk, start, end3

total68,096

What the other side of the budget costs

Москва6 characters13

مرحبا5 characters11

北京2 characters7

🙏1 character5

Scripts outside the eleven get no learned tokens at all — they fall through to raw bytes at two to four tokens per character.

The allocation is the design decision. Devanagari takes roughly double the next script, which tracks Hindi being about a fifth of the training corpus; Latin still takes the largest single share. Every slot given to one script is a slot withheld from another — which is why this one bar explains both the fertility win and its limits, with no benchmark involved.

Read from the published Sarvam-1 tokenizer.json·@sushant_p18