Compare / DeepSeek V3.2 vs gemini-3-flash
DeepSeek V3.2 vs gemini-3-flash
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 | DeepSeek V3.2 | gemini-3-flash |
|---|---|---|---|
| Arena (formerly LMArena)reported | OverallArena rating | 1,437 | 1,468 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,423 | 1,469 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,400 | 1,457 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,453 | 1,496 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,420 | 1,458 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,425 | 1,473 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,410 | 1,461 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 21.5 | 26.3 |
| Artificial Analysisreported | GPQA Diamond% of questions | 84.0% | 89.8% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 24.6% | 36.6% |
| Artificial Analysisreported | IFBench% of instructions | 60.7% | 78.0% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 73.3% | 78.0% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 35.6% | 38.6% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 90.6% | 80.4% |
| OpenHands Indexreported | Average% resolved | 45.7% | 49.0% |
| OpenHands Indexreported | Front end (SWE-Bench Multimodal)% resolved | 27.9% | 22.1% |
| OpenHands Indexreported | Greenfield (Commit0)% resolved | 25.0% | 18.8% |
| OpenHands Indexreported | Information gathering (GAIA)% resolved | 50.3% | 58.8% |
| OpenHands Indexreported | Issue resolution (SWE-Bench)% resolved | 71.6% | 74.6% |
| OpenHands Indexreported | Testing (SWT-Bench)% resolved | 53.6% | 70.7% |
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.