Skip to content
Leaderboard

GaMS3-12B-Instruct

CJVTopen weights

Licence
Gemma Terms of Use
ID
cjvt/GaMS3-12B-Instruct

GaMS3-12B-Instruct has a rating of 1,060 (±55) in the Slovenian Chat Arena, set by 392 votes in blind comparisons in Slovenian. The rating's 95% interval is 1,005–1,115: it shows how precisely the votes have pinned the rating down so far. It wins 61.2% of its decided comparisons (172 wins, 109 losses, 111 ties). The median length of its answers is 201 words. It is developed by CJVT. Its weights are open (licence: Gemma Terms of Use).

At a glance

Rating
1,060±55
95% interval 1,005–1,115
Answer length
201words
median, typically 33–525
Votes
392
61.2% of decided votes won

Model information

GaMS (Generative Model for Slovene) by the Centre for Language Resources and Technologies of the University of Ljubljana (CJVT). Instruction model built on Google Gemma 3 12B.

More about the model (opens in a new tab)

Statistics

Rating with 95% confidence interval

  1. GaMS2-27B-Instruct
  2. GaMS-27B-Instruct-Nemotroninactive
  3. gpt-5
  4. GaMS3-12B-Instruct
  5. gemini-2.0-flashinactive
  6. Mistral-Large-3-675B-Instruct-2512inactive
  7. gemma-3-27b-itinactive

The models just above and below it on the leaderboard, on one axis. The upright mark is the rating, the line its 95% interval.

Vote outcomes

  • Wins172
  • Ties111
  • Losses109
Total queries
601

Answer length among models

Median words per answer for every model on the leaderboard, on one axis.

Longer answers than 24 of 39 other models.

All models in numbers
ModelMedian words
gpt-oss-120b463
Ministral-3-14B-Instruct-2512348
deepseek-r1336
gemini-2.5-pro330
Qwen3-Next-80B-A3B-Instruct312
gemma-4-26b-a4b-it306
Qwen3.6-27B284
gpt-oss-20b284
gemini-3-pro-preview260
gemini-2.5-flash242
GaMS-9B-Instruct-Nemotron229
GaMS-27B-Instruct-Nemotron226
gemini-3.1-pro-preview224
Mistral-Large-3-675B-Instruct-2512224
deepseek-r1-0528216
GaMS3-12B-Instruct201
nemotron-3-ultra-550b-a55b199
GaMS2-27B-Instruct196
gemma-4-31b-it179
gemini-3.8-flash178
llama3-70b-instruct168
gpt-5-mini152
kimi-k2-instruct142
gemini-3.5-flash141
Mistral-Small-24B-Instruct-2501110
gemma-3-12b-it107
bielik-11b-v2.3-instruct104
gpt-5101
gemma-3-27b-it101
gemini-1.5-pro100
llama-4-scout-17b-16e-instruct91
llama-3.3-70b-instruct90
EuroLLM-9B-Instruct83
gemini-2.0-flash78
llama-4-maverick-17b-128e-instruct78
gpt-5.6-luna77
gpt-4o68
Zlatorog-12B-Instruct-Beta47
GaMS-9B-Instruct13
GaMS-27B-Instruct10

Answer length in tokens

Share of answers by token count, from 56 answers.

A token is a piece of a word as the model counts it. The model's reasoning is not counted.

Data by bin
TokensAnswersShare
0–1281323%
128–256916%
256–38435%
384–51235%
512–64035%
640–76835%
768–89624%
896–1,02459%
1,024–1,15235%
1,152–1,280611%
1,280–1,40800%
1,408–1,53635%
1,536–1,66400%
1,664–1,79200%
1,792–1,92024%
1,920–2,04800%
2,048–2,17612%
2,176–2,30400%
2,304–2,43200%
2,432–2,56000%
2,560–2,68800%
2,688–2,81600%
2,816–2,94400%
2,944–3,07200%
3,072–3,20000%
3,200–3,32800%
3,328–3,45600%
3,456–3,58400%
3,584–3,71200%
3,712–3,84000%
3,840–3,96800%
3,968–4,09600%
4,096–4,22400%
4,224–4,35200%
4,352–4,48000%
4,480–4,60800%
4,608–4,73600%
4,736–4,86400%
4,864–4,99200%
4,992–5,12000%
5,120–5,24800%
5,248–5,37600%
5,376–5,50400%
5,504–5,63200%
5,632–5,76000%
5,760–5,88800%
5,888–6,01600%
6,016–6,14400%
6,144–6,27200%
6,272–6,40000%
6,400–6,52800%
6,528–6,65600%
6,656–6,78400%
6,784–6,91200%
6,912–7,04000%
7,040–7,16800%
7,168–7,29600%
7,296–7,42400%
7,424–7,55200%
7,552–7,68000%
7,680–7,80800%
7,808–7,93600%
7,936–8,06400%
8,064–8,19200%