vlm.fyi Get the data
Models / MoonshotAI

Kimi K3

Measured on 3 providers. The model card lists text, image, video input; what each provider accepts is below.

Fireworks

accounts/fireworks/models/kimi-k3
Images
120 “Too many images were provided, we currently limit the number of images per conversation to 120”
768×432 frames
112 “Too many images were provided, we currently limit the number of images per conversation to 120”
Tokens / frame
457
Inline body
≥ 32 MB
Images by URL
80 MB “Image <url> download size exceeded maximum size limit of 10.0 MB”
Video file
96 MB “Video <url> download size exceeded maximum size limit of 100.0 MB”
Video tok/min
? “no complete response within 600s”
Frames/min
—
Video input
Yes
Tool messages
Lenient
Extra fields
Rejects usage, reasoning
  • Fireworks counts every image in the messages array ("per conversation"), not per message.

Measured 2026-10-31

Modal

moonshotai/Kimi-K3
Images
≥ 1,024
768×432 frames
≥ 384
Tokens / frame
457
Inline body
≥ 32 MB
Images by URL
≥ 128 MB
Video file
No video “Kimi-K3 supports image input only”
Video tok/min
No “Kimi-K3 supports image input only”
Frames/min
—
Video input
No
Tool messages
Lenient
Extra fields
All accepted

Measured 2026-10-31

Together AI

moonshotai/Kimi-K3
Images
≥ 512?
768×432 frames
≥ 384
Tokens / frame
457
Inline body
≥ 32 MB
Images by URL
≥ 64 MB?
Video file
≥ 512 MB
Video tok/min
Ignored
Frames/min
—
Video input
Ignored
Tool messages
Strict order
Extra fields
All accepted
  • Accepts a video_url with HTTP 200, but the video adds no tokens: the model never sees it.
  • Together caps the decoded media in one request at ≈1.43 GB (seen on DeepSeek V4.1 Flash; likely shared).

Measured 2026-10-31