Compare / Qwen3 235B A22B Thinking 2507 vs Gemma 4 26B A4B
Qwen3 235B A22B Thinking 2507 vs Gemma 4 26B A4B
20 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 Thinking 2507 | Gemma 4 26B A4B |
|---|---|---|---|
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,404 | 1,444 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,370 | 1,399 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,466 | 1,475 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,385 | 1,437 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,400 | 1,438 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,377 | 1,406 |
| Artificial Analysisreported | Artificial Analysis Coding Indexindex score | 22.1 | 39.3 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 12.7 | 16.7 |
| Artificial Analysisreported | GPQA Diamond% of questions | 79.0% | 79.2% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 15.9% | 19.3% |
| Artificial Analysisreported | IFBench% of instructions | 51.2% | 72.4% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 72.0% | 65.7% |
| Artificial Analysisreported | SciCode% of problems | 41.4% | 40.0% |
| Artificial Analysisreported | Terminal-Bench 2.1% of tasks | 12.0% | 39.0% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 13.6% | 25.0% |
| Artificial Analysisreported | Τ-bench banking% of tasks | 7.8% | 12.0% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 53.2% | 43.6% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 30.0% | 8.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.31 | 0.35 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 32.4 | 43.8 |
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.