Compare / Qwen3.5 397B A17B vs DeepSeek V3.1 Terminus
Qwen3.5 397B A17B vs DeepSeek V3.1 Terminus
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 | Qwen3.5 397B A17B | DeepSeek V3.1 Terminus |
|---|---|---|---|
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,445 | 1,423 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,405 | 1,405 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,434 | 1,420 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,442 | 1,417 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,421 | 1,411 |
| Artificial Analysisreported | Artificial Analysis Coding Indexindex score | 48.2 | 43.5 |
| Artificial Analysisreported | Artificial Analysis Intelligence Indexindex score | 21.4 | 14.8 |
| Artificial Analysisreported | GPQA Diamond% of questions | 89.3% | 79.2% |
| Artificial Analysisreported | Humanity's Last Exam% of questions | 29.0% | 16.4% |
| Artificial Analysisreported | IFBench% of instructions | 78.8% | 57.0% |
| Artificial Analysisreported | Long-context reasoning (AA-LCR)% of questions | 77.3% | 69.3% |
| Artificial Analysisreported | SciCode% of problems | 44.8% | 38.0% |
| Artificial Analysisreported | Terminal-Bench 2.1% of tasks | 51.3% | 44.9% |
| Artificial Analysisreported | Terminal-Bench 4.0% of tasks | 0.0% | 0.0% |
| Artificial Analysisreported | Terminal-Bench Hard% of tasks | 40.9% | 31.8% |
| Artificial Analysisreported | Τ-bench banking% of tasks | 13.4% | 21.0% |
| Artificial Analysisreported | Τ²-bench telecom% of tasks | 95.6% | 37.1% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 23.0% | 32.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.33 | 0.37 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 50.1 | 54.0 |
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.