#14AMay 9, 2026Evidence A

Keye-VL-2.0

Name: Keye-VL-2.0 multimodal review
Item: Keye-VL-2.0
Rating: 86
Author: CN Multimodal Tier List

Kuaishou · 30B-A3B (MoE)

The open specialist for long-video timing

First to bring DSA sparse attention to multimodal understanding; near-lossless long-video timing at 256K. Real short-video platform data is its training moat.

TextImageVideoBolt-on

Last updated Oct 18, 2026Editorial scoring

Bottom line

For video-only work, Keye-VL-2.0 is the first open model to try: cheap, temporally accurate, easy to deploy. OCR and generation are clearly weaker.

Eight-dimension breakdown

editorial mapping · 0–100

Keye-VL-2.086.0

Visual perception & grounding20%

Document parsing & Chinese OCR18%

Chart & document reasoning14%

Video & long-horizon reasoning14%

GUI agents & tool use14%

Generation & editing6%

Chinese context & culture6%

Ecosystem, cost & availability8%

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
上下文	256K	DSA sparse attention
激活参数	3B	30B total

Sources:

Keye-VL 开源仓库快手 KwaiVGI

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		84 · Standard documents only
Chart data extraction		83 · Simple charts
30-min video QA		95 · Standout; near-lossless timing
Mobile GUI control		81 · Average
Chinese cultural context		85 · Strong in short-video context

Strengths and limitations

Strengths

256K long-video timing with accurate event localization.
Low 30B-A3B cost suits batch video processing.
Open for fine-tuning with strong short-video domain data.

Limitations

Document and chart quality are not top tier.
Weak generation; unusable for design.
Average Chinese cultural common sense.

Price and real cost

List price

¥2/M tokens

Open; API ≈ ¥2 / 1M output. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥1.2 (assuming input at 25% of output price).

Use your volume

About Keye-VL-2.0

Is Keye-VL-2.0 open source?

Licence: Open (community license). Weights are available, but check the exact terms and filing requirements before commercial use.

How much does Keye-VL-2.0 cost?

Open; API ≈ ¥2 / 1M output. It ranks #14 with a composite of 86.0. Real bills depend on resolution, tiling, caching and volume discounts.

What is Keye-VL-2.0 best for?

Long-video moderation; Video chaptering and highlight localization; Batch video tagging。

Why does Keye-VL-2.0 score 86.0?

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 Video & long-horizon reasoning; 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.