Compare / Qwen3 235B A22B Instruct 2507 vs Gemma 3 12B
Qwen3 235B A22B Instruct 2507 vs Gemma 3 12B
17 results from 3 sources that measured both models. Each row is in the source's own units; there is no overall winner.
| Benchmark | Metric | Qwen3 235B A22B Instruct 2507 | Gemma 3 12B |
|---|---|---|---|
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,432 | 1,386 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,378 | 1,332 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,448 | 1,274 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,414 | 1,321 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,422 | 1,342 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,396 | 1,313 |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Irrelevance detection% correct | 81.7% | 70.3% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Memory% correct | 23.9% | 27.5% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Multi-turn tasks% correct | 45.4% | 5.8% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Overall accuracy% correct | 52.1% | 30.4% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Relevance detection% correct | 93.8% | 93.8% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Single-turn calls (curated)% correct | 90.3% | 79.4% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Single-turn calls (user-contributed)% correct | 78.7% | 74.2% |
| Berkeley Function Calling Leaderboard (BFCL) V4reported | Web search% correct | 54.0% | 4.0% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 15.0% | 9.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.36 | 0.29 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 49.0 | 29.9 |
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.