Intern-S1: benchmark results
Scientific multimodal model built on a 235B Qwen3 MoE language model and a 6B InternViT vision encoder. Further pretrained on 5T multimodal tokens, including over 2.5T scientific-domain tokens; this is additional training, not the full foundation-model budget. Provider: Shanghai AI Lab. Released 2025-07-27. Access: Open.
Aggregate Rating 1663 ± 13, rank #285 of 1656 rated models, from 67 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| V2X-QA - Cooperative Average | 43.5 | Accuracy (%) on the cooperative view (vehicle and roadside i | 100 |
| V2X-QA - Cooperative Visibility | 70.2 | Accuracy (%) on the cooperative Visibility task (CO2); four- | 100 |
| V2X-QA - RSU Traffic Agents | 61.5 | Accuracy (%) on the infrastructure-side RSU Traffic Agents t | 100 |
| SGI-Bench Deep Research | 15.74 | Deep Research Score | 94.4 |
| V2X-QA - Infrastructure-Side Average | 59.7 | Accuracy (%) on the infrastructure-side (roadside unit) view | 94.4 |
| MMMU Benchmark | 77.7 | Validation Score | 91.1 |
| V2X-QA - Ego Visible Agents | 63.8 | Accuracy (%) on the vehicle-side Ego Visible Agents task (VS | 88.9 |
| V2X-QA - RSU Global Layout | 82.3 | Accuracy (%) on the infrastructure-side RSU Global Layout ta | 88.9 |
| V2X-QA - RSU Global Risk | 60.8 | Accuracy (%) on the infrastructure-side RSU Global Risk task | 88.9 |
| V2X-QA - Vehicle-Side Average | 58.5 | Accuracy (%) on the vehicle-side (ego camera) view, mean of | 88.9 |
| V2X-QA - Ego Risk | 42.3 | Accuracy (%) on the vehicle-side Ego Risk task (VS3); four-o | 77.8 |
| V2X-QA - Ego Static Scene | 83.8 | Accuracy (%) on the vehicle-side Ego Static Scene task (VS1) | 77.8 |
Interactive version: theaggregate.ai/model?slug=intern-s1 · How It Works · Data refreshed daily, snapshot 2026-10-20.