Qwen·released 2024-09-191 source
Qwen2.5-72B benchmark scores: 18 benchmarks tracked. 17% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 18 benchmarks·best result 92.67 on ARC AI2 (self-reported)·3 of 18 independently reproduced
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| ARC AI2 | 92.67accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| MMLU | 80.4accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| HellaSwag | 79.73accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| BBH | 73.07accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| TriviaQA | 71.9accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| PIQA | 65.2accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| Winogrande | 64.6accuracy (%) | self-reported· optimizedT1 | 2024-09-19DeepSeek-V3 Technical Report |
| MATH level 5 | 63.17accuracy (%) | reproduced· optimizedT1 | 2024-09-19Epoch AI |
| GeoBench | 62accuracy (%) | unverifiedT1 | 2024-09-19GeoBench leaderboard |
| DTBench | 38.22accuracy (%) | unverifiedT1 | 2024-09-19https://conceptualreasoning.ai/dtbench |
| METR Time Horizons | 35.78accuracy (%) | unverifiedT1 | 2024-09-19METR - Measuring AI Ability to Complete Long Tasks |
| GPQA diamond | 32.2accuracy (%) | reproduced· optimizedT1 | 2024-09-19Epoch AI |
| Balrog | 16.2accuracy (%) | unverifiedT1 | 2024-09-19https://balrogai.com/ |
| WeirdML | 15.97accuracy (%) | unverifiedT1 | 2024-09-19WeirdML Leaderboard |
| LMCA | 15.79accuracy (%) | unverifiedT1 | 2024-09-19https://conceptualreasoning.ai/lmca |
| OTIS Mock AIME 2024-2025 | 7.96accuracy (%) | reproduced· optimizedT1 | 2024-09-19Epoch AI |
| The Agent Company | 5.7accuracy (%) | unverifiedT1 | 2024-09-19TheAgentCompany experiment results github |
| OSWorld | 5accuracy (%) | unverified· optimizedT1 | 2024-09-19OS World Website |
Claims drawn from cited facts, not live model generation.
Developed by Qwen and released in 2024, this model is a large-scale system with 72 billion parameters, a size that implies greater capability headroom alongside higher computational and deployment costs than smaller alternatives.
3 cited facts
This model is scored on 18 tracked benchmarks, but it holds no current top score on any of them. Its record can be trusted: the results have been independently reproduced, meaning outside evaluation backs the scores up rather than resting on vendor claims alone.
3 cited facts
On average this model trails the state-of-the-art leader by 37.44 points, a score gap under whatever disclosed harness produced it rather than a capability claim — and a gap that wide sits in the well-behind-the-leaders band, not the couple-of-points frontier band or a merely moderate competitive one. Consistent with that, it holds the top score on none of its benchmarks, which means it has no genuine category leadership to point to yet.
3 cited facts
2 facts cross-checked across data sources: 2 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: curated_param_counts, epoch_benchmarks
Qwen2.5-72B is an AI model developed by Qwen, released 2024-09-19. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Qwen2.5-72B has recorded scores on 18 benchmarks, each shown with its evidence status.
Qwen2.5-72B has recorded scores on 18 benchmarks — PIQA, Winogrande, GPQA diamond, MATH level 5, OTIS Mock AIME 2024-2025, TriviaQA, and 12 more. The full table above shows each score with its evidence status.
3 of 18 recorded scores (17%) are independently reproduced rather than self-reported by the lab.
Qwen2.5-72B has 18 tracked claims: 3 independently reproduced, 7 self-reported, 8 unverified.