InternVL3-78B: benchmark results
Shanghai AI Lab's open InternVL3 78B vision-language model. Provider: Shanghai AI Lab. Released 2025-04-11. Access: Open.
Aggregate Rating 1631 ± 5, rank #352 of 1656 rated models, from 1419 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CausalPhys | 59.24 | Answer accuracy (0-1 scaled to %) over all four domains (Per | 100 |
| CausalPhys - Description Correctness | 47.2 | Description Correctness (0-1 scaled to %), the share of attr | 100 |
| CausalPhys - Entity Faithfulness | 69.68 | Entity Faithfulness (0-1 scaled to %), the share of causal-g | 100 |
| CausalPhys - Perception | 62.57 | Answer accuracy (0-1 scaled to %) on the Perception domain ( | 100 |
| CausalPhys - Relation Awareness | 30.52 | Relation Awareness (0-1 scaled to %), the share of directed | 100 |
| FREAK (Multiple Choice) - Consistency Accuracy | 35.9 | Consistency accuracy (%): share of FREAK's 1,000 multiple-ch | 100 |
| Flat-Pack Bench - Temporal Ordering | 43.87 | TOrd: order the connection events of the highlighted parts, | 100 |
| MEGA-Bench Task - COCO Person Detection | 79.9 | Task Score (%) | 100 |
| MEGA-Bench Task - Insect Order Classification | 46.7 | Task Score (%) | 100 |
| MEGA-Bench Task - Perception Test Video Action Count | 75 | Task Score (%) | 100 |
| MEGA-Bench Task - Stack Overflow Debug QA | 71.4 | Task Score (%) | 100 |
| MEGA-Bench Task - VLN Identify Location | 57.6 | Task Score (%) | 100 |
Interactive version: theaggregate.ai/model?slug=internvl3-78b · How It Works · Data refreshed daily, snapshot 2026-10-20.