Compare / Nova 2 Lite vs Gemma 3 27B

Nova 2 Lite vs Gemma 3 27B

20 results from 4 sources that measured both models. Each row is in the source's own units; there is no overall winner.

Nova 2 Lite
Amazon · profile
Gemma 3 27B
Google · profile
Results compared
20
Shared sources
4
BenchmarkMetricNova 2 LiteGemma 3 27B
Arena (formerly LMArena)reportedOverallArena rating 1,3801,380
Arena (formerly LMArena)reportedBusiness, management and financeArena rating 1,3511,377
Arena (formerly LMArena)reportedCreative writingArena rating 1,2691,348
Arena (formerly LMArena)reportedExpert promptsArena rating 1,3541,337
Arena (formerly LMArena)reportedInstruction followingArena rating 1,3271,343
Arena (formerly LMArena)reportedOverallArena rating 1,3351,365
Arena (formerly LMArena)reportedWriting, literature and languageArena rating 1,2981,354
Berkeley Function Calling Leaderboard (BFCL) V4reportedIrrelevance detection% correct 82.1%73.7%
Berkeley Function Calling Leaderboard (BFCL) V4reportedMemory% correct 2.4%13.6%
Berkeley Function Calling Leaderboard (BFCL) V4reportedMulti-turn tasks% correct 2.1%10.8%
Berkeley Function Calling Leaderboard (BFCL) V4reportedOverall accuracy% correct 27.1%29.5%
Berkeley Function Calling Leaderboard (BFCL) V4reportedRelevance detection% correct 75.0%81.2%
Berkeley Function Calling Leaderboard (BFCL) V4reportedSingle-turn calls (curated)% correct 87.0%87.2%
Berkeley Function Calling Leaderboard (BFCL) V4reportedSingle-turn calls (user-contributed)% correct 80.8%74.5%
Berkeley Function Calling Leaderboard (BFCL) V4reportedWeb search% correct 5.0%0.0%
UGI LeaderboardreportedRequested-length error% off the requested word count 45.0%8.0%
UGI LeaderboardreportedStyle adherencescore from 0 to 1 0.350.29
UGI LeaderboardreportedWriting scorescore out of 100 23.545.0
Vectara Hallucination LeaderboardreportedAnswer rate% of documents 99.6%98.8%
Vectara Hallucination LeaderboardreportedHallucination rate% of summaries 5.1%7.4%

Bold green marks the better value on that metric. Where a source reports ranges that overlap, the difference may not be meaningful; see the benchmark page for ranges.