InternVL3-8B: benchmark results
OpenGVLab/Shanghai AI Lab's open 8B vision-language model (April 2025), natively multimodal-pretrained with tool-use and GUI-agent capabilities. Provider: Shanghai AI Lab. Released 2025-04-11. Access: Open.
Aggregate Rating 1547 ± 4, rank #620 of 1656 rated models, from 1943 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CrashSight | 68.7 | Accuracy (%) on all seven question categories (the paper's A | 100 |
| CrashSight - Crash Mechanics | 61.1 | Accuracy (%) on crash-mechanics questions (collision type, i | 100 |
| CrashSight - Fault Determination | 71.3 | Accuracy (%) on fault-determination questions (primary cause | 100 |
| CrashSight - Involved Parties | 58.1 | Accuracy (%) on involved-party questions (vehicle and vulner | 100 |
| CrashSight - Post-Crash Outcome | 80.3 | Accuracy (%) on post-crash outcome questions (road-user stat | 100 |
| CrashSight - Robustness | 82.2 | Accuracy (%) on hallucination-probe questions (cannot-determ | 100 |
| CrashSight - Temporal Sequence | 85.7 | Accuracy (%) on temporal-sequence questions (event order and | 100 |
| FewMMBench (Zero-Shot) | 82.6 | Mean accuracy (%) on the FewMMBench query sets (about 250 mu | 100 |
| GeoDrive-Bench | 61 | Overall accuracy (%) on 5,053 human-validated multiple-choic | 100 |
| MEGA-Bench Task - Deciphering Oracle Bone | 21.4 | Task Score (%) | 100 |
| MEGA-Bench Task - GUI Act Web Single | 13.1 | Task Score (%) | 100 |
| MEGA-Bench Task - Logical Reasoning 2D Folding | 50 | Task Score (%) | 100 |
Interactive version: theaggregate.ai/model?slug=internvl3-8b · How It Works · Data refreshed daily, snapshot 2026-10-20.