InternVL3.5-4B: benchmark results
Provider: Shanghai AI Lab. Released 2025-08-26. Access: Open.
Aggregate Rating 1527 ± 9, rank #1985 of 10916 rated models, from 427 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EgoArgus (Assistance Decision) - Video-Grounded On-Topic | 85.71 | Assistance-decision F1 (%; whether help is required, from th | 100 |
| NutriBench-Kitchen - Ingredient Entry (Hard) | 33.3 | Score (%; the ingredient entry task family, hard regime: dif | 100 |
| UAV-DualCog - Flight Behavior Recognition (Composite) | 39.8 | Composite-level behavior accuracy (%; 1,024 first-person UAV | 96.2 |
| RoadSafe365 - Dense Captioning | 0.32 | METEOR (0-1) of the model's detailed description of each acc | 95 |
| MVX-Bench - Defeasible Entailment | 53.3 | Accuracy (%) on the 150 Defeasible Entailment questions (dec | 94.7 |
| RoadSafe365 - Accident/Violation | 47.91 | Four-option multiple-choice accuracy (%) on the accident or | 90.9 |
| UXBench (Mobile) - Content Mismatch | 68 | Accuracy (0-100) on service names inconsistent with the page | 86.4 |
| EgoCoT-Bench - Goal-Oriented Object Tracking | 38.21 | Answer accuracy (%, four-way multiple choice, strict exact m | 83.3 |
| MultihopSpatial - 1-Hop Ego-Centric | 32.8 | Multiple-choice accuracy (%) on the 1-hop ego-centric questi | 80.6 |
| ChronoBench - Land Cover Perception | 52.03 | Accuracy (%; level average of object perception, the land co | 80 |
| MVX-Bench - Sequence Reasoning | 44.4 | Accuracy (%) on the 117 Sequence Reasoning questions (order | 78.9 |
| MIOH - Attribute | 38 | Accuracy (%; attribute questions about object attributes acr | 77.8 |
Interactive version: theaggregate.ai/model?slug=internvl3-5-4b · How It Works · Data refreshed daily, snapshot 2026-11-08.