Head-to-head
Intern-S2-Preview-397B vs Qwen3.8-27B
A 6.1-point composite gap — but the real story is the dimension mix. Same weights, unpacked: who leads where, and whether the premium is justified.
#26 / 56
Intern-S2-Preview-397B
Shanghai AI Laboratory
A new open base for scientific multimodal work
Full dossierThe short answer
The composites differ by 6.1 points, and their strengths do not overlap.
Stronger overall: Qwen3.8-27B. If you only run one workload, read the dimension table instead.
Overlaid radar
Same weights, same scale; overlap means parity
- Intern-S2-Preview-397B84.4
- Qwen3.8-27B90.5
Dimension deltas
A on the left, B on the right; 100 max per row
Intern-S2-Preview-397B leads 2, Qwen3.8-27B leads 6.
Spec sheet
Hard specs in one table
| Item | Intern-S2-Preview-397B | Qwen3.8-27B |
|---|---|---|
| Composite | 84.4 | 90.5 |
| Tier | A- | A+ |
| Visual perception & grounding | 92 | 94WIN |
| Document parsing & Chinese OCR | 91 | 94WIN |
| Chart & document reasoning | 92WIN | 91 |
| Video & long-horizon reasoning | 72 | 88WIN |
| GUI agents & tool use | 76 | 88WIN |
| Generation & editing | 60 | 80WIN |
| Chinese context & culture | 87 | 92WIN |
| Ecosystem, cost & availability | 90WIN | 88 |
| Output price | ¥0 / M | ¥1 / M |
| Context | 256K | 262K (up to 1M) |
| Parameters | 397B total (MoE) | 27B dense |
| Licence | Open weights (Apache-2.0) | Open weights (Apache-2.0) |
| Native multimodal | Yes | Yes |
| Evidence grade | A | A |
How to choose
Choose Intern-S2-Preview-397B if…
- Research papers and experimental charts
- Domain image-text work in materials, biology and remote sensing
- Domain continued pretraining under Apache-2.0
- You cannot accept its limitation: Video 72 and generation 60: motion content and creative generation are out of scope.
How to choose
Choose Qwen3.8-27B if…
- Single-machine or small-cluster on-prem
- Domain document fine-tuning
- Cost-controlled image-text batch work
- You cannot accept its limitation: Generation 80 and video 88: creative work and long video still need a larger model.
Scenario check
Five scenarios, who holds up
| Scenario | Intern-S2-Preview-397B | Qwen3.8-27B | Better |
|---|---|---|---|
| Research chart extraction | 92 | 94 | Qwen3.8-27B |
| Paper layout parsing | 91 | 91 | Tie |
| 30-min video QA | 70 | 86 | Qwen3.8-27B |
| Mobile GUI control | 72 | 86 | Qwen3.8-27B |
| Chinese academic context | 88 | 92 | Qwen3.8-27B |
Comparison FAQ
Intern-S2-Preview-397B vs Qwen3.8-27B: which is stronger?
Composite: Intern-S2-Preview-397B 84.4, Qwen3.8-27B 90.5. Intern-S2-Preview-397B leads on Chart & document reasoning, Ecosystem, cost & availability; Qwen3.8-27B leads on Visual perception & grounding, Document parsing & Chinese OCR, Video & long-horizon reasoning, GUI agents & tool use, Generation & editing, Chinese context & culture. The stronger overall is Qwen3.8-27B.
How much more expensive is Intern-S2-Preview-397B than Qwen3.8-27B?
Intern-S2-Preview-397B is ¥0/1M output tokens; Qwen3.8-27B is ¥1/1M. On output price, Qwen3.8-27B is pricier; real cost also depends on image tokens and input share.
Can Qwen3.8-27B replace Intern-S2-Preview-397B?
It depends on your primary workload. If it sits in Qwen3.8-27B’s strengths (Single-machine or small-cluster on-prem, Domain document fine-tuning, Cost-controlled image-text batch work), yes. If you need Research papers and experimental charts, switching will cost quality.
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.