InternVL 4.0
Shanghai AI Laboratory · Undisclosed (MoE, ~9B active)
The open-source perception ceiling, with video as the priced-in trade
Almost absent from public comparisons, yet a long-time top-three download in the open-source community. Perception 96 is the highest open-weights score below the closed flagships, MIT licence, official API at ¥1/1M output tokens; the price is images and text only — video 66 is third from the bottom here. It is the only model on this 26-model board where community heat and ranking heat run in opposite directions.
Last updated Oct 18, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- InternVL 4.085.0
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 |
|---|---|---|
| OpenCompass 多模态榜 | 78.9 | #1 open-weights entry (official) |
| MMMU | 81.2 | Official, same prompts for val/test |
| 输出单价 | ¥1 | Official API; weights free |
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 |
|---|---|---|
| Chinese receipt OCR | 93 · Strong on print, mid on handwriting | |
| Chart data extraction | 90 · Reliable on single figures | |
| 30-min video QA | 62 · Frame-level description only | |
| Mobile GUI control | 74 · Not trained for GUI | |
| Fine-grained recognition | 96 · Same as our perception dimension |
Strengths and limitations
Strengths
- Perception 96 — the highest open-weights score, behind only Doubao and Qwen3.8-Max at 97.
- MIT licence plus an official API: switching between self-hosted and managed carries no legal cost.
- Grade A: the most-reproduced open model here, with weights and eval scripts fully public.
Limitations
- Images and text only: no video, audio or GUI modality, and agent 78 sits below the enterprise floor.
- Video 66 is third from the bottom here — frame-level description at best, no temporal reasoning.
- Ecosystem 84: the official API trails the Qwen family on stability and docs, and self-hosting means building your own inference stack.
Price and real cost
List price
¥1/M tokens
≈ ¥1 / 1M output tokens (official API); weights free. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥0.6 (assuming input at 25% of output price).
Use your volumeAbout InternVL 4.0
Is InternVL 4.0 open source?
Licence: Open weights (MIT). Weights are available, but check the exact terms and filing requirements before commercial use.
How much does InternVL 4.0 cost?
≈ ¥1 / 1M output tokens (official API); weights free. It ranks #16 with a composite of 85.0. Real bills depend on resolution, tiling, caching and volume discounts.
What is InternVL 4.0 best for?
Fine-grained recognition and product-attribute extraction; Commercial on-prem work that needs an MIT licence; A perception component inside an open-source stack。
Why does InternVL 4.0 score 85.0?
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 Video & long-horizon reasoning.
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.