Intern-S2-Preview-397B
Shanghai AI Laboratory · 397B total (MoE)
A new open base for scientific multimodal work
A 397B scientific multimodal base released in July 2026 under Apache-2.0: charts 92, documents 91, perception 92 and a composite of 84.4. It replaces the sixth edition’s unverifiable InternVL 4.0 entry, which has no matching version in the official repository or papers. S2 is strong on research figures and paper layouts, and clearly weak elsewhere: video 72, agents 76 and generation 60, so it is not a general multimodal workhorse.
Last updated Oct 21, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- Intern-S2-Preview-397B84.4
Weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8. Definitions and sources on the methodology page.
Public benchmarks
as reported, not converted| Benchmark / metric | Value | Note |
|---|---|---|
| 参数规模 | 397B | Official model card |
| 许可证 | Apache-2.0 | Commercial use and modification |
| 上下文 | 256K | Official spec |
Public scores come from different versions and prompts and are not interchangeable with our composite; they are one evidence input.
Scenario mapping
five stress tests, one scale| Scenario | Score | Observation |
|---|---|---|
| Research chart extraction | 92 · Stable on multi-panel figures and axes | |
| Paper layout parsing | 91 · Reads formulas and two-column layouts | |
| 30-min video QA | 70 · Sparse-frame understanding only | |
| Mobile GUI control | 72 · Not trained for GUI | |
| Chinese academic context | 88 · Accurate terminology; average on colloquial text |
Strengths and limitations
Strengths
- Charts 92 and documents 91: among the best open-weight scores for research figures and paper layouts.
- Apache-2.0 plus 397B: clean licensing and enough capacity for domain continued pretraining.
- Grade A: weights, model card and evaluation scripts are all public and reproducible.
Limitations
- Video 72 and generation 60: motion content and creative generation are out of scope.
- Agents 76 is below the enterprise floor; tool use needs an outer orchestration layer.
- Self-hosting 397B is expensive; small teams are better served by a hosted API or a distilled model.
Price and real cost
List price
—N/A
Open weights; self-hosted, no API fee. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥0.0 (assuming input at 25% of output price).
Use your volumeAbout Intern-S2-Preview-397B
Is Intern-S2-Preview-397B open source?
Licence: Open weights (Apache-2.0). Weights are available, but check the exact terms and filing requirements before commercial use.
How much does Intern-S2-Preview-397B cost?
Open weights; self-hosted, no API fee. It ranks #26 with a composite of 84.4. Real bills depend on resolution, tiling, caching and volume discounts.
What is Intern-S2-Preview-397B best for?
Research papers and experimental charts; Domain image-text work in materials, biology and remote sensing; Domain continued pretraining under Apache-2.0。
Why does Intern-S2-Preview-397B score 84.4?
The composite weights eight dimensions: perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. Its strongest dimension is Visual perception & grounding; its weakest is Generation & editing.
Personal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.