Kimi K3
Moonshot AI · 2.8T total (MoE, open)
The engineer’s bet: long context plus native vision
K3 brings native vision into the main model: CharXiv 84.8%, MathVision 94.3%, plus 2M context — built for reading a whole manual and its screenshots together.
Last updated Oct 18, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- Kimi K392.3
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 |
|---|---|---|
| CharXiv | 84.8 | Chart understanding |
| MathVision | 94.3 | Visual math reasoning |
| 上下文 | 2M | 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 |
|---|---|---|
| Chinese receipt OCR | 90 · Clean docs fine; handwriting average | |
| Chart data extraction | 95 · Excellent on paper-grade charts | |
| 30-min video QA | 88 · Usable, not best-in-class | |
| Mobile GUI control | 94 · Strong cross-app planning | |
| Chinese cultural context | 92 · Factual; weaker on memes |
Strengths and limitations
Strengths
- 2M context plus native vision reads a whole manual in one pass.
- Strong chart and math visual reasoning for paper reproduction.
- Open weights enable deep fine-tuning and on-prem use.
Limitations
- Weak generation/editing; not for creative design pipelines.
- Self-hosting 2.8T weights is impractical; open here means auditable.
- Slow at peak; long-context requests are costly.
Price and real cost
List price
¥16/M tokens
≈ ¥16 / 1M output tokens. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥9.7 (assuming input at 25% of output price).
Use your volumeAbout Kimi K3
Is Kimi K3 open source?
Licence: Open weights (community license). Weights are available, but check the exact terms and filing requirements before commercial use.
How much does Kimi K3 cost?
≈ ¥16 / 1M output tokens. It ranks #5 with a composite of 92.3. Real bills depend on resolution, tiling, caching and volume discounts.
What is Kimi K3 best for?
Engineering manuals and API docs QA; Paper-figure reproduction and code screenshots; Very long multimodal knowledge bases。
Why does Kimi K3 score 92.3?
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.